Keywords: Cyber-physical systems and security, Robotics and swarm intelligence, Control theories
Abstract: Distributed formation control is a fundamental functionality of multi-robot systems, enabling autonomous vehicles to maneuver cohesively without relying on centralized infrastructure. However, achieving robust performance in the presence of relative measurement mismatches, parametric uncertainties, and heterogeneous sensor topologies, while safely navigating obstacle-strewn environments, remains a major barrier to real-world adoption. This plenary talk presents recent theoretical and practical advances aimed at closing these critical gaps.

First, we introduce adaptive and dynamical control laws designed to eliminate measurement mismatches, compensate for parametric uncertainties, and reject external disturbances. We then delve into the challenges of operating with heterogeneous sensor systems and present approaches to mitigate these problems. Extending this work to use a low-cost sensing solution, we examine the use of vision-based relative information to achieve the formation. To guarantee safe transit, we discuss the integration of distributed Control Barrier Functions directly into the formation control design. Finally, we introduce a paradigm-shifting formation control problem by moving beyond rigid geometric shapes to focus on how a swarm can dynamically achieve a desired spatial distribution function across a given space.

Keywords: Control theories, Nonlinear control and applications, Autonomous vehicles
Abstract: This paper considers safe path-following control for a vehicle robot using a zeroing control barrier function (ZCBF). Previous studies have proposed path-following control that makes a vehicle robot follow a reference point moving along a predefined path, as well as control barrier function-based methods for safe control. In addition, trajectory-tracking control considering obstacle avoidance has been studied for two-wheeled vehicle robots using control barrier function. However, in trajectory-tracking, delays due to obstacle avoidance increased tracking errors, leading to excessive control inputs. Therefore, in this study, we employ a path-following control that can avoid the aforementioned problems.

Based on these studies, this research designs a safe control law for a vehicle robot that combines an existing path-following controller and a zeroing control barrier function. However, simply combining the two is insufficient, so we propose a new path-following control input that also takes obstacle avoidance into consideration. Finally, simulations confirm that the vehicle robot can avoid an obstacle while following the reference path.

Keywords: System identification and modelling, Industrial applications, Nonlinear control and applications
Abstract: This paper presents a framework for model identification and control system design to improve the robustness of a spinbath circulation process. A grey-box dynamic model is developed using historical steady-state data, and three PID control configurations are evaluated: single-loop, conventional cascade, and summed-setpoint cascade controllers. Simulation results are presented to show that the summed-setpoint cascade control configuration provides better performance, reducing peak error by 61% and IAE by 68% compared with the existing single-loop and conventional cascade control configuration.

Keywords: Control theories, Cyber-physical systems and security, Industrial applications
Abstract: The evolution of modern power systems (PSs) has increased their susceptibility to cyberattacks, potentially leading to performance deterioration or system destabilization. In PSs, load frequency control (LFC) is essential for maintaining frequency stability; however, its effectiveness is challenged by cyber threats. Therefore, this study introduces a novel triple-parametric fractional controller (TPFC) strategy that integrates the benefits of proportional-integral (PI) and proportional-derivative (PD) methods, thereby bypassing the need for external control loops and improving system efficacy. We use the stability boundary locus (SBL) approach to determine the optimal controller configurations. The TPFC’s effectiveness is demonstrated through simulation studies that alleviate random and step-load disruptions amid multiple cyber threats in microgrids, outperforming existing methods while preserving frequency stability. Moreover, the performance evaluation using integral errors clearly shows the efficacy of the proposed strategy.

Keywords: Control theories, Deep learning and machine learning, System identification and modelling
Abstract: This paper addresses a unified framework combining the robustness guarantees of robust linear quadratic regulator (RLQR) designs with the exploration capabilities of maximum entropy (ME) reinforcement learning (RL). The RLQR problem is merged with ME control problem with a modified cost functional that explicitly accounts for parametric uncertainty. The key concept is to use entropy regularization to promote exploration in the control policy, while simultaneously guaranteeing robustness against bounded parameter variations. The resulting controller is characterized by a modified Riccati equation and yields an optimal policy that is Gaussian-distributed, with its mean corresponding to the RLQR control and covariance determined by the entropy regularization parameter. Connections to model-free off-policy RL are discussed and a comparative evaluation with standard off-policy algorithm is carried out thereby validating improved performance of the proposed approach compared to standard methods.

Keywords: Control theories, Industrial applications
Abstract: Most industrial multivariable processes often have multiple time delays and significant interactions. For large-scale multivariable processes, the severity of these effects is even greater. As a result, for higher-order multivariable processes, designing an independent controller for each paired input-output (i-o) loop is complex. Eliminating deadtime dynamics and reducing the impact of interaction effects can facilitate the design of a promising controller for multivariable processes. To achieve this, deadtime compensation (DTC) based multiloop configurations are established in this work. The proposed diagonal DTC (DDTC) and full DTC (FDTC) based multiloop control frameworks can make the process transfer function (PTF) matrix elements free from time delays. Such elimination of the long transport delays from the process dynamics can explicitly reduce the impact of the interaction effect among the paired i-o loops. Subsequently, it simplifies the parametrization of an independent controller for each paired i-o loop that is a perfect control approximation (equivalent transfer function (ETF)). IMC-PI tuning rules are utilized to evaluate controller parameters for corresponding ETF models. Further, it has been mathematically proven that the error in each paired i-o channel is independent of the transport delays with the proposed DTC-based control schemes. To validate the efficacy of the proposed control approaches, a simulation study was carried out on popular multivariable syst

Keywords: Control theories, Intelligent control, Computational intelligence
Abstract: This paper investigates a distributed Nash equilibrium (NE) seeking problem in an online game, where each agent minimizes its time-varying cost function subject to its own feasible set. The cost function may depend on the decisions of other agents, which is only revealed after each agent commits its decision during the game. The agents can communicate exclusively with their immediate neighbors over an undirected connected network. However, as the problem scale increases, communication bottlenecks inevitably arise from the substantial communication overhead of exchanging real-valued information. To address this challenge, a quantized distributed online NE seeking algorithm is proposed, integrating projected gradient descent with a leaderfollowing consensus protocol. A compression quantizer is utilized to quantize the estimation errors exchanged among agents. The algorithm’s performance is assessed by an expected dynamic regret, which is proven to grow sublinearly under mild assumptions. Simulation results validate the theoretical findings.

Keywords: Control theories, Robotics and swarm intelligence, Nonlinear control and applications
Abstract: This paper studies disturbed multi-agent source seeking under hard safety constraints. Instead of tracking prescribed trajectories or rigid templates, the agents minimize a composite source-seeking, cohesion, and Fisher-information cost that induces a task-driven sensing geometry. To separate disturbance estimation from safety filtering, we combine an executed-input disturbance observer with a pre-QP residual gate and a robust CBF-QP safety filter. The gate schedules both observer injection and robust margins using nominal-input safety residuals, thereby avoiding algebraic loops and mitigating the disturbance-observer CBF (DOB-CBF) conflict. We prove forward invariance of the safety set and free-region ISS tube convergence to the unconstrained optimal set. Simulations demonstrate information-shaped emergent formations and constrained safe passage with reduced post-passage conservatism.

Keywords: Control theories, Nonlinear control and applications, Health systems
Abstract: This paper proposes an interval type-2 Mamdani fuzzy control framework for a six-compartment cancer treatment model involving CD4 T cells, CD8 T cells, sensitive cancer cells, resistant cancer cells, dendritic cells, and IL-2. The model includes three therapeutic modalities, namely chemotherapy, radiotherapy, and immunotherapy. To regulate these treatments, three biological indicators are introduced, namely tumor burden, resistance level, and immune condition, which serve as input to the fuzzy controller. In addition, a therapy scheduler is incorporated to prevent unrealistic simultaneous administration of all treatments over the entire 2100 hours. Numerical simulations show that, without control, the immune related compartments decline while the cancer populations eventually regrow. In contrast, under the proposed fuzzy control with therapy scheduling, both sensitive and resistant cancer cells are effectively suppressed, while the immune-related compartments remain positive and exhibit treatment-driven responses. These results indicate that the proposed approach provides an effective and interpretable strategy for multimodal cancer therapy under biological uncertainty.

Keywords: Artificial intelligence, Autonomous vehicles, Deep learning and machine learning
Abstract: Large Vision-Language Models(LVLMs) are increasingly deployed in real-world systems that require reliable multimodal decision. However, recent studies have shown that misleading textual cues embedded in visual scenes can significantly disrupt semantic reasoning. Existing approaches primarily focus defenses on perception level, often overlook how such semantic disturbances propagate through the reasoning process, leading to unstable or unreliable outputs. We propose a Robust Perception-Aware Decision framework (RPAD) that improves decision robustness under semantic disturbances. Instead of directly eliminating adversarial signals, RPAD models typographic attacks as semantic perturbations and introduces a confidence-guided modality regulation mechanism to dynamically balance visual and textual contributions. In addition, a decision grounding strategy is employed to enforce consistency between generated outputs and verified visual evidence, reducing reliance on corrupted semantic cues. Extensive experiments on multiple benchmarks demonstrate that RPAD significantly improves decision robustness across different LVLMs, while maintaining competitive performance on clean data.

Keywords: Artificial intelligence, Big data, Computational intelligence
Abstract: Multivariate time series (MTS) anomaly detection in complex cyber-physical systems is crucial for ensuring operational safety. However, accurate detection remains immensely challenging due to dynamic spatiotemporal dependencies and the irregular sampling nature of real-world sensor data. Most existing graphbased methods rely on discrete-time modeling, which frequently suffers from temporal information loss under non-uniform time intervals. To overcome these limitations, this paper proposes a novel end-to-end framework, namely Correlation-aware Dynamic Anomaly Detection (CoD-AD). CoD-AD leverages a Dynamic Correlation Learning module to adaptively capture time-varying, asymmetric spatial topologies. Building upon this, a Continuously Controlled Evolution module utilizes Graph Neural Controlled Differential Equations to map discrete observations into a continuous control path, explicitly capturing continuous temporal dynamics. Furthermore, a robust Principal Component Analysis scoring mechanism is integrated to isolate genuine anomalous signals from operational noise and achieve precise root cause diagnosis via topological contribution aggregation. Extensive experiments on real-world SWaT and WADI datasets under irregular sampling conditions demonstrate that CoD-AD significantly outperforms state-of-the-art baselines in detection accuracy, early warning capability, and root cause localization.

Keywords: Artificial intelligence, Big data, Control devices, sensors and actuators
Abstract: This paper presents a Software-Defined Vehicle (SDV) architecture for special-purpose electric vehicles, extending the SDV concept beyond traditional applications. The proposed platform adopts a centralized software-based vehicle architecture that integrates vehicle functions, application management, and Over-The-Air (OTA) updates into a unified system. A modular SDV-based vehicle platform is designed to support multiple heterogeneous special-purpose functions, including road sweeping, aerial ladder operation, road ice removal, and on-site additive manufacturing (3D printing), which can be dynamically reconfigured through software applications on a single vehicle. To ensure safe operation across different functional modes, a hybrid AI-based fault management framework is introduced. An LSTM model trained on healthy data predicts expected system behavior, while deviations between predicted and observed signals are treated as anomaly indicators. These indicators are evaluated by an LLM-based rule-driven reasoning module to perform fault detection, confidence estimation, and decision stabilization. A confidence-gated fault-latching mechanism is applied to ensure stable driver alerts, and OTA actions are triggered only after fault confirmation, without reboot or rollback. The results demonstrate that the proposed framework enables stable, explainable, and safety-oriented fault detection , supporting scalable and reliable SDV-based special-purpose vehicle operation.

Keywords: Artificial intelligence, Control devices, sensors and actuators, Health systems
Abstract: Mechanical ventilation is a supportive therapy for patients with respiratory failure, both inadequate and prolonged ventilation pose significant health risks, including increased mortality. However, the optimal duration of mechanical ventilation is difficult to predict and relies heavily on real-time clinical assessment. To address the challenge, we propose a novel predictive algorithm termed RLMV. Specifically, we represent the first attempt to integrate static physiological features with dynamic temporal variations, constructing a cascaded prediction model that combines "Feature Regression Screening" with "Temporal Deep Modeling". Experimental results based on the MIMIC-III dataset demonstrate the superiority of RLMV, which achieving a Mean Squared Error (MSE) of 0.0032 and a Mean Absolute Error (MAE) of 0.0324 and outperforming the existing methods.

Keywords: Artificial intelligence, Control devices, sensors and actuators, Health systems
Abstract: This study proposes an AI-assisted smartwatch-based emergency monitoring framework that integrates multi-level activity intensity estimation with apnea-like event detection. To move beyond conventional categorical human activity recognition, activities are reformulated into five ordinal intensity levels, enabling continuous modeling of motion dynamics. A novel dataset, textit{Dyn-Intensity}, was constructed to capture high-intensity and emergency-like wrist motions. Lightweight features derived from Signal Vector Magnitude (SVM) and its temporal variation enable efficient on-device computation. In parallel, infrared PPG signals are analyzed for apnea-like detection using subject-wise baseline normalization and logistic regression. Experimental results demonstrate that MLP-based intensity estimation achieves high ordinal agreement (QWK = 97.27%), while apnea-like detection attains an F1-score of 91.8% under subject-independent evaluation. The proposed framework demonstrates the feasibility of real-time, resource-efficient emergency monitoring on wearable devices.

Keywords: Artificial intelligence, Deep learning and machine learning, Industrial applications
Abstract: The fiber optic gyroscope (FOG), as the core of an inertial navigation system, is highly sensitive to temperature fluctuations, which degrade navigation accuracy. Conventional temperature compensation methods, developed under stable conditions, fail to address errors caused by dynamic thermal gradients and component drifts. This paper presents an AI-driven temperature compensation approach based on an extended LSTM (xLSTM) network. The method establishes an adaptive sequence modeling framework that captures multi-step temporal dependencies in FOG measurements. By incorporating multi-level feature fusion and a memory-based architecture, the model effectively handles complex thermal dynamics. Joint optimization of instantaneous and time-interval compensation losses further enhances accuracy and long-term stability. Experimental results demonstrate that the proposed xLSTM significantly improves temperature compensation performance, offering a practical solution for the miniaturization and intelligent enhancement of high-precision autonomous navigation systems.

Keywords: Artificial intelligence, Deep learning and machine learning, Industrial applications
Abstract: This study presents an artificial intelligence (AI)-based approach to predict kerosene yield from condensate using its physical and chemical properties. Accurate yield prediction is critical for optimizing refinery operations and improving product quality, supporting sustainable and low-carbon energy production. Conventional methods relying on manual analysis or traditional statistical models often suffer from lower accuracy, inconsistency, and longer processing times. To overcome these limitations, an XGBoost predictive model was developed using a dataset comprising density, specific gravity, kinematic viscosity, API gravity, sulphur content, and total acid number. The model was trained and tested to fine-tune performance, achieving a high coefficient of determination (R² = 0.9732) and demonstrating superior capability in capturing complex relationships compared to linear regression, artificial neural networks (ANN), and LightGBM. This AI-driven methodology offers a systematic, efficient, and sustainable tool for refinery decision-making, supporting cleaner and more optimized energy production.

Keywords: Artificial intelligence, Deep learning and machine learning, Intelligent control
Abstract: Speech-based assistive navigation offers a natural interaction modality for users with severe motor impairments. However, speech variability in cerebral palsy (CP) significantly degrades the performance of static recognition models. This study investigates session-wise incremental adaptation for speech-based assistive navigation under limited data conditions. A VGG16-based recognition model is first trained on healthy speech data and evaluated on CP speech, where baseline accuracy drops from approximately 0.94 to 0.44–0.60 across sessions. To address this gap, a hybrid incremental learning strategy is introduced that integrates progressive parameter unfreezing, replay-based stabilization, and adaptive learning rate control. Experimental results show that the proposed strategy achieves a session-by-session maximum accuracy of 0.92 while maintaining strong retention in early CP sessions (0.85–0.89). These findings demonstrate that controlled incremental adaptation effectively balances plasticity and stability for assistive navigation systems operating under non-stationary and user-specific speech conditions.

Keywords: Artificial intelligence, Deep learning and machine learning
Abstract: Landslide detection is an essential task in geological hazard monitoring, which is critical for disaster prevention and emergency response. Unmanned Aerial Vehicles (UAVs) offer flexible, high-resolution monitoring capabilities, and the acquired RGB optical images and Digital Elevation Model (DEM) provide complementary information. This paper proposes a UAV-based cross-modal object detection method for landslides. First, the cross-modal detection process is formulated, and the characteristics of the two modalities are analyzed. Then, the cross-modal fusion network is designed, incorporating high-frequency and lowfrequency components to adaptively fuse features of texture and structure. Furthermore, an improved YOLOv13 detection model with a regression stabilization gate module is established to better capture landslide features. Experimental results on UAV landslide datasets verify that the proposed method outperforms both single-modality and other fusion methods in landslide detection effectiveness and accuracy.

Keywords: Industrial applications, Measurement and instrumentation
Abstract: This paper presents a computer vision system for automated deflection monitoring of the blast furnace downcomer using unmanned aerial vehicle (UAV) imagery. The method isolates the pipe region via color-based segmentation, establishes a reference line via polynomial fitting, and extracts the pipe centerline through skeletonization and cubic curve fitting. Furthermore, perspective distortion is corrected using a perspective transformation model solved via singular value decomposition, thereby ensuring precise restoration of physical scale. Experimental results show the system effectively handles complex industrial backgrounds, corrects imaging distortions, and provides reliable quantitative deflection measurements and visualizations, significantly improving inspection safety and efficiency.

Keywords: Control devices, sensors and actuators, Nonlinear control and applications, System identification and modelling
Abstract: Thermo-responsive hydrogel actuator (TRHA) is pivotal for soft robotics due to their biological muscle-like movement. However, their practical application is hindered by complex energy conversion mechanisms, asymmetric hysteresis, and rate-dependent hysteresis nonlinearities. This paper proposes an inverse dynamic compensation method to achieve the high-precision tracking control. The inner loop employs a proportional-integral (PI) controller to ensure stable thermal response and eliminate environmental disturbances. The outer loop utilizes a composite control method that combines an inverse compensation feedforward controller based on a modified Prandtl-Ishlinskii (P-I) model and a polynomial model with a PI feedback to effectively mitigate asymmetric hysteresis and rate-dependent hysteresis. Furthermore, two experiments are conducted to validate the effectiveness of the controller. The root mean square errors for the two experiments are 3.45% and 4.79%, respectively. These results demonstrate that the method provides a robust and high-precision solution for deploying hydrogel-based soft robots in complex environments.

Keywords: Artificial intelligence, Deep learning and machine learning
Abstract: Humans generally expect robots to dynamically recognize and understand the communication atmosphere, while two key challenges remain unaddressed well in Human-Robot Interaction (HRI). The first is that due to the complexity and fuzziness of communication atmospheres, robots struggle to model them dynamically. The second is that in real life, emotionally rich interaction backgrounds interfere with robots' perception and modeling of atmospheres. To address these challenges, we propose a Multi-hierarchical Fuzzy Communication Atmosphere Dynamic modeling method (MFCAD) and a Multimodal Emotion Modeling Human-Robot Interaction system (MEM-HRI). Considering the richness of emotions embedded in music, we use it as the background of the interaction. Specifically, in MFCAD, built upon Fuzzy Atmosfield (FA) and Fuzzy Analytical Hierarchy Process (FAHP), we employ a three-layer hierarchical framework (an atmosphere layer, an emotion layer, and an influence layer) to dynamically model atmospheres and design SVM-SVR Networks (SSN) to recognize emotions of the background. In MEC-HRI, robots can dynamically feel and model atmospheres with MFCAD and enhance human-robot communication experiences by adjusting background music, robot modes, and actions. HRI experiments are performed with 6 volunteers, two Pepper robots, and three Nao robots. Four pieces of music from the DEAM database accompany HRI in different scenarios.

Keywords: Motion and vibration control, System identification and modelling, Intelligent control
Abstract: Liquid metal droplets exhibit unique fluidic and metallic properties, making them attractive for applications such as soft robotics and microfluidics. However, existing control methods typically rely on fixed electrode configurations, which significantly constrain the accessible workspace. To address this limitation, this paper proposes a composite electrokinematic control framework for two-dimensional path tracking of a liquid metal droplet using a single pair of mobile electrodes. The electrode pair, mounted on a robotic arm, is modeled as an electric dipole whose position can be actively adjusted to dynamically regulate the electric field distribution. The proposed controller integrates electrode positioning based on an electric dipole model with voltage regulation via a PI controller. For practical implementation, a target-switching mechanism with error compensation is incorporated to ensure continuous path-following performance. Experimental results demonstrate the effectiveness of the proposed approach.

Keywords: Industrial applications, Deep learning and machine learning, Computational intelligence
Abstract: The static weighing of trains requires that the carriages be precisely parked within the effective measurement area. However, traditional weighing methods basedon personnel guidance is inefficient, while other weighing strategiesutilizing various sensors solutions are prone to environmental interference. To address these challenges, this paper proposes a visual positioning method based on data augmentation. Firstly, to deal with complex lighting changes, a contrast-limited adaptive histogram equalization preprocessing strategy in the CIELAB space is adopted to decouple the brightness and enhance the local contrast. Then, the YOLOv13n detector is introduced, and the HyperACE mechanism based on hypergraph is utilized to capture the global semantic correlation between couplers and wheels, ensuring robust detection under complex conditions. Additionally, a 5th-order polynomial regression model is established to map pixel coordinates to physical distances, eliminating the need for complex calibration to correct the distortion of wide-angle lenses. Application results in a certain steel enterprise show that the average precision mean (mAP@0.5) of this method exceeds 99%, and the inference delay is only 1.97 ms. The mean absolute error (MAE) of physical positioning is 3.2 cm, fully meeting the accuracy and real-time requirements for automated static weighing.

Keywords: Robotics and swarm intelligence, Mechatronics, Measurement and instrumentation
Abstract: Aiming at the problems of noise pollution and poor environmental adaptability of traditional propeller-driven underwater vehicles, we design a fish-tail-like soft robot (FTLSR) based on the fish’s Body/Caudal Fin (BCF) propulsion mode. The actuator adopts an antagonistic drive structure composed of a polypropylene (PP) bionic fishbone and asymmetric pneumatic bionic silicone muscles, realizing underwater movement by simulating fish tail swinging. To reveal its motion characteristics, three experimental platforms are established to investigate the relationships between driving pressure, tail oscillation displacement, static thrust and turning torque. In addition, tail trajectory tracking control experiments are carried out using a PID controller, verifying the feasibility of swing propulsion. This work lays a foundation for the engineering application of the FTLSR. Future work will focus on modeling based on these characteristic relationships to achieve autonomous underwater swimming control.

Keywords: Mechatronics, Autonomous vehicles, Control devices, sensors and actuators
Abstract: The objective of this study is to enable mobile robots operating in environments shared with pedestrians to move safely and efficiently without obstructing pedestrians. Generally, robots travelling on pavements detect the pedestrian flow directly in front of them and move along that flow, thereby reaching their destination without obstructing pedestrians. However, in environments where pedestrians are densely concentrated, such as parks, university campuses and theme parks, it is reasonable for the robot not only to avoid the pedestrian flow directly in front of it, but also to select areas with fewer pedestrians within the environment from a large-area perspective. Therefore, this study evaluates the difficulty of the robot’s movement based on environmental perception from a bird’s-eye view obtained from a drone flying above the robot (this study defines this as 'passability'). Furthermore, this study utilizes the passability to design navigation routes that enable the robot to move safely and efficiently without obstructing pedestrians, while adapting to the constantly changing pedestrian movement. The effectiveness of our proposed method was confirmed through simulation experiments using video data captured by an actual drone.

Keywords: Mechatronics, Autonomous vehicles, Control devices, sensors and actuators
Abstract: The objective of this study is to enable mobile robots to move safely and efficiently without obstructing pedestrians by predicting their movement intentions. Conventional avoidance methods based on physical observations, such as position and velocity, often fail to respond effectively to sudden changes in pedestrian behavior because pedestrians are defined simply as moving objects. On the other hand, pedestrians have habits unique to humans, such as turning their faces in the direction they are walking. Therefore, this study focuses on the “pedestrian’s facial yaw angle” and the "walking direction vector". By utilizing the characteristics that humans turn their faces before changing direction, we propose a method to detect the early signs of course changes. Specifically, we utilize YOLOv8, DeepSort, and an LSTM network to predict pedestrians’ future heading directions. The effectiveness of the proposed method is verified through experiments using an actual mobile robot, demonstrating a 19% reduction in arrival time.

Keywords: Computational intelligence, Deep learning and machine learning
Abstract: Single-photon lidar imaging is critically impaired in dense fog due to strong backscattering from microscopic water droplets, which substantially increases scattered photon counts and degrades the effective signal-to-noise ratio. This leads to severe challenges in accurately extracting target returns, resulting in significant depth estimation bias and frequent detection errors. However, traditional signal processing methods, including temporal and frequency-domain filtering and parametric estimation, often fail under high-density fog due to poor signal separability. To overcome these limitations, we combine Monte Carlo photon propagation with Mie scattering and exploit the temporal distribution differences between target echo photons and fog-induced backscattered photons to generate realistic single-photon lidar datasets under varying fog densities. Based on this dataset, we introduce a Transformer-Depth model that employs self-attention mechanisms to jointly model temporal photon sequences and spatial features, effectively capturing long-range dependencies and enhancing temporal features of the photon signal. Extensive evaluations on both simulated and real-world foggy scenarios demonstrate that the proposed approach significantly improves depth estimation accuracy and robustness, outperforming conventional physics-based methods and temporal network models.

Keywords: Control theories, System identification and modelling, Intelligent control
Abstract: We propose a differential evolution (DE) algorithm specifically tailored for the design of linear-quadratic-Gaussian (LQG) controllers in quantum systems. Building upon the foundational DE framework, the algorithm incorporates specialized modules, including relaxed feasibility rules, a scheduled penalty function, adaptive search range adjustment, and the “bet-and-run” initialization strategy. These enhancements improve the algorithm’s exploration and exploitation capabilities while addressing the unique physical realizability requirements of quantum systems. The proposed method is applied to a quantum optical system, where three distinct controllers with varying configurations relative to the plant are designed. The resulting controllers demonstrate superior performance, achieving lower LQG performance indices compared to existing approaches. In addition, the algorithm ensures that the designs comply with physical realizability constraints, guaranteeing compatibility with practical quantum platforms. The proposed approach holds significant potential for application to other linear quantum systems in performance optimization tasks subject to physically feasible constraints.

Keywords: Deep learning and machine learning, Industrial applications, Measurement and instrumentation
Abstract: Accurate status monitoring of transmission lines is fundamental to ensuring the safety of the power grid. However, ``apparent tilt'' caused by undulating terrain and sensor acquisition angles complicates the detection of genuine tower tilting, subsidence, and conductor sag, necessitating high-fidelity 3D reconstruction. In foggy environments, traditional optical sensors are limited by low visibility and signal attenuation, whereas single-photon detection (SPD) emerges as a vital solution due to its extreme sensitivity. To address the challenges of high signal-to-background ratios (SBR), structural fragmentation, and the lack of annotated SPD data in fog, this paper conducts research on two levels. First, based on the SPD mechanism, a physical simulation model is developed using Gamma and Gaussian distributions for fog and target modeling, respectively. High-fidelity datasets under various visibility conditions are generated via the Monte Carlo method. Second, a noise-guided 3D reconstruction algorithm is proposed. Leveraging the powerful feature extraction capabilities of Generative Adversarial Networks (GANs), the algorithm achieves progressive and refined reconstruction of power facilities directly from intense background noise. The results demonstrate that the proposed method effectively restores the essential spatial poses of facilities, providing reliable data support for power safety early warning.

Keywords: System identification and modelling
Abstract: The estimation of all the parameters in an unknown quantum state or measurement device, commonly known as quantum state tomography (QST) and quantum detector tomography (QDT), is crucial for comprehensively characterizing and controlling quantum systems. In this paper, we introduce a framework, in two different bases, that utilizes multiple quantum processes to simultaneously identify a quantum state and a detector. We develop a closed-form algorithm for this purpose and prove that the mean squared error (MSE) scales as O(1/N) for both QST and QDT, where N denotes the total number of state copies. Furthermore, we formulate the problem as a sum of squares (SOS) optimization problem with semialgebraic constraints, where the physical constraints of the state and detector are characterized by polynomial equalities and inequalities.

Keywords: Control theories
Abstract: An augmented system model provides an effective way to model non-Markovian quantum systems, which is useful in filtering and control for this class of systems. However, since a large number of ancillary quantum oscillators representing internal modes of a non-Markovian environment directly interact with the principal system in these models, the dimension of the augmented system may be very large causing signiffcant computational burden in designing filters and controllers. In this context, this paper proposes an H2 model reduction method for the augmented model of linear non-Markovian quantum systems. We ffrst establish necessary and sufffcient conditions for the physical realizability of the augmented model of linear non-Markovian quantum systems, which are more stringent than those for Markovian quantum systems. However, these physical realizability conditions of augmented system model pose non-convex constraints in the optimization problem of model reduction, which makes the problem different from the corresponding classical model reduction problem. To solve the problem, we derive necessary conditions for determining the input matrix in the reduced model, with which a theorem for designing the system matrix of the ancillary system in the reduced system is proved. Building on this, we convert the nonlinear equality constraints into inequality constraints so that a semideffnite programming algorithm can be developed to solve the optimization problem for model reduction.

Keywords: Control theories, Nonlinear control and applications
Abstract: We summarize a constructive framework for ensemble control of driftless two-level quantum systems. Building upon our previous work on explicit pulse-design method based on interaction-frame reformulation and Fourier synthesis, we emphasize two developments. First, the same design philosophy extends to driftless n-level ensembles via decomposition into embedded two-level channels of mathfrak{su}(n). Second, a Magnus-based refinement improves the accuracy/switching tradeoff: a third-order correction with sign alternation upgrades the error bound from O(N^{-1}+Nvarepsilon^{p}) to O(N^{-3}+Nvarepsilon^{p}). Hence, comparable precision can be achieved with fewer switching segments. We also outline how this framework can be combined with drift-compensation techniques.

Keywords: Artificial intelligence, Deep learning and machine learning, Computational intelligence
Abstract: Hyperspectral image (HSI) classification requires modeling complex spectral--spatial dependencies under high dimensionality and limited supervision. Most existing methods rely on Euclidean feature representations and generic feature mixing, which may be insufficient for explicitly capturing high-order cross-band correlations. This paper presents QUEEN, a quantum-inspired embedding and evolution network for HSI classification. QUEEN formulates representation learning as a unified Hilbert-space pipeline including quantum-state embedding, factorized unitary evolution, joint projective measurement, and fidelity-guided optimization. Multi-scale HSI patches are encoded as quantum-inspired states, transformed through structured unitary evolution, and fused via measurements on a composite Hilbert space. Experiments on public benchmarks show that QUEEN achieves competitive or superior performance, while offering a physically interpretable perspective for structured dependency modeling in hyperspectral analysis.

Keywords: System identification and modelling, Control theories
Abstract: Hamiltonian estimation is an effective approach in studying the structure and dynamical evolution of quantum systems. The difficulty in estimating the Hamiltonian is that an N-qubit Hamiltonian has 4^N-1 unknown parameters, requiring exponentially many equations for information extraction. In this paper we develop a method based on compressed sensing to estimate the Hamiltonian of a multi-qubit system. We identify a problem where as N increases, the common sufficient condition (Restricted Isometry Property) for compressed sensing often fails, obstructing the application of compressed sensing in (N>= 3)-qubit Hamiltonian estimation. To solve this problem, we propose a ``scale transformation" technique to restore RIP and ensure a compressive estimation of a k-sparse Hamiltonian using only O(klog(4^N/k)) equations. In the numerical examples, we estimate the Hamiltonians of a 30-qubit systems, and obtain the relative error lower than 20%, demonstrating the effectiveness of the method.

Keywords: System identification and modelling
Abstract: This paper investigates estimation of the values and higher-order derivatives of transfer function matrix (TFM) at a prescribed set of points in the complex plane for continuous-time descriptor systems. An extended steady-state output decomposition is developed for the case where the input generating system has multiple eigenvalues, which enables an explicit separation of TFM values and their derivatives at different eigenvalues in steady-state response. Furthermore, we analyze the informativity of the data and derive sufficient conditions under which the estimation matrix is of full row rank, thereby ensuring unique estimation of the TFM values and derivatives. Numerical simulations are provided to validate the theoretical results.

Keywords: System identification and modelling, Computational intelligence
Abstract: We extend the Wide-Array of Nonlinear Dynamics Approximations (WyNDA) framework to enable data-driven discovery of partial differential equations (PDEs). The proposed approach fuses an adaptive-observer scheme with WyNDA’s model identification to jointly reconstruct latent states and uncover the governing PDE structure and coefficients. We demonstrate the method on two canonical benchmarks: the linear heat equation and the nonlinear Burgers’ equation. In both cases, the identified coefficients converge rapidly to their true values, and the reconstructed states closely match reference solutions, indicating accurate dynamical recovery with modest data requirements. These results suggest that WyNDA, augmented with adaptive observation, offers a practical and effective foundation for discovering physical laws in complex distributed-parameter systems.

Keywords: System identification and modelling, Control theories, Control devices, sensors and actuators
Abstract: This paper presents a control algorithm for autonomous resonance tracking in electrothermal-actuated resonators. Existing resonance-tracking approaches either assume that the resonator exhibits second-order dynamics or enforce oscillations at a frequency that may not coincide with the resonator’s true resonance. The electrothermal actuated resonators inherently exhibit third-order dynamics, making conventional methods unsuitable. The proposed technique, inspired by phase-locked loop principles, employs a driving signal composed of three related frequencies and exploits the corresponding phase-lag information to achieve autonomous frequency tracking. As a result, the resonator can operate at a frequency of maximum sensitivity without requiring prior knowledge of resonator parameters. Simulation results show that, with a 10 ms integration interval and an SNR of 6.02, the method attains a frequency lock-on error of 0.08%. This algorithm is implemented on a digital controller, and hardware-in-the-loop verification is performed using an equivalent plant consisting of a second-order RLC circuit cascaded with a first-order RC circuit. Experimental results indicate lock-on errors ranging from 0.34% to 3.86%, depending on the separation between the electrothermal and mechanical characteristic frequencies.

Keywords: Control theories, System identification and modelling
Abstract: This study proposes a data-driven inverse optimal control method for discrete-time linear time-invariant systems that allows time-varying quadratic cost weights to accommodate changing design intent. Under the assumption of blockwise constant weights, the inverse problem can be reduced to a linear homogeneous equation, and an identifiability condition is also derived in terms of the rank of the resulting coefficient matrix. Numerical examples on a double-integrator system and an inverted pendulum system demonstrate that the proposed method estimates time-varying weights more accurately than the conventional method based on the time-invariant assumption. These results illustrate the effectiveness of the proposed framework for recovering time-varying design intent directly from input-output data.

Keywords: System identification and modelling, Industrial applications
Abstract: Recycle processes are often characterized by significant delays that lead to staircase-like transient responses and pose significant challenges in model identification and control. Traditional data-driven modeling approaches struggle to identify these class of processes due to complex model representations and associated challenges. This work proposes a parsimonious data-driven modeling framework for identifying recycle dynamics with dominant delays. A two-step identification procedure is developed: first, the structure of discrete-time model capable of explaining the recycle dynamics is identified using ARX models combined with block orthogonal matching pursuit (BOMP); second, the structural insights are incorporated to carry out continuous-time parameter estimation of the model using prediction error minimization (PEM). Monte Carlo simulations are carried out to demonstrate the efficacy of proposed model identification approach.

Keywords: System identification and modelling, Industrial applications, Deep learning and machine learning
Abstract: This paper explores using instance segmentation to improve inspections of marine growth on mooring lines, aiming to reduce costs and labor by reducing the mooring cleaning campaigns. The approach proposed in this paper helps maintenance decision-making. The research begins by exploring available object detection and segmentation methods. Instance segmentation is chosen as it provides pixel-level identification. A dataset has been developed, and YOLOv5 and YOLOv8 models are employed to detect and estimate the thickness level of marine growth. The results are combined with a depth-detection algorithm to produce depth profiles. Evaluation metrics such as F1 score, Precision, Recall, and Confusion Matrix are used, whose results indicate that YOLOv8s-seg is a promising candidate for automated marine growth inspection under the conditions considered in this study.

Keywords: System identification and modelling, Mechatronics, Intelligent control
Abstract: Identification of nonlinear mechanical systems using a closed-loop measurement is hard since it suffers from control errors, lack of experimental excitation, and the inaccuracy of data-based models in representing real components. This paper introduces a physics informed closed-loop system identification framework that estimates plant parameters and controller gains jointly through closed-loop measurements. The system under consideration is a differentiable rotary inverted pendulum for which dynamic model includes learnable motor and controller parameters, and a bounded residual form. In this paper, PD controllers for balancing and swing-up energy shaping of the considered inverted pendulum are developed, which are specifically included in the learning loop and the soft hybrid gate to ensure a smooth transition in the controller. One-step prediction, open-loop and closed-rollout, control imitation and MAP priors are some of the drivers of learning function. The predictions with sparse data for the Quanser QNET 2.0 prove to be accurate when it is experimentally validated using five excitation trajectories. The resulting model has less than 5000 trainable parameters and has a reference inference latency of less than 5 µs, which is sufficient to run physical digital twins on embedded assemblages and edge-control workloads.

Keywords: System identification and modelling, Robotics and swarm intelligence, Autonomous vehicles
Abstract: The real-time application of a swarm of quadrotors is safety-critical. As they operate within a network, a fault in a single agent can propagate to others, potentially leading to total system failure. Therefore, a faulty agent must be quickly isolated to avoid further damage. Existing literature demonstrates that faulty agents can be effectively managed if the fault characteristics are known. However, in real-world scenarios, we typically only have access to input-output data from the quadrotors to determine their onboard status. To address it, this paper adopts a data-driven Koopman framework for the online estimation of actuator faults in quadrotors. A distributed averaging algorithm is employed to reach a swarm-wide consensus on fault status, allowing for the isolation of faulty agents and dynamic reformation of the remaining swarm. The effectiveness of this architecture is validated through numerical simulations of a five-quadrotor swarm performing predefined formations.

Keywords: Nonlinear control and applications, Control theories, Medical and financial systems
Abstract: We study finite-horizon stochastic control issues for affine systems driven by a Poisson random measure, with controls entering linearly through the drift. We prove a bang-bang-like principle: for a broad class of suitable terminal criteria, optimizing over the closed convex hull of the control set yields the same value as optimizing over strict controls. Furthermore, optimal controls can be approximated by policies switching between extremal values. We then apply this result to a two-line Cramér-Lundberg insurance model with common Poisson claims and equity transfers, and use Monte Carlo simulations of bang-bang transfer strategies to show how the initial allocation of a fixed total equity and the premium loading determine whether internal transfers can keep both lines solvent at the horizon.

Keywords: Nonlinear control and applications, Control theories, Medical and financial systems
Abstract: We study a continuous-time rational addiction model where addiction capital follows a piecewise deterministic Markov process with state-dependent jumps capturing relapse and recovery. The instantaneous utility combines consumption and addiction capital via a power specification, leading to Hamilton-Jacobi-Bellman (HJB) equations with nonlocal jump terms. In the capped, bounded-control case we obtain a unique bounded viscosity solution and prove that optimal policies are bang-bang, switching between minimal and maximal consumption, while in the uncapped case we derive explicit linear feedback controls and closed-form value functions in several parameter regimes.

Keywords: Control theories
Abstract: In this paper, we study mean-field backward doubly stochastic differential equations (BDSDEs) and their optimal control problems, where the coefficients depend on both the state of the solution process and its expected value. The first part of the paper is devoted to establish the existence and uniqueness of solutions for such mean-field BDSDEs under monotonicity conditions. Furthermore, under the same monotonicity assumptions, we derive Pontryagin's maximum principle, which serves as a necessary condition for optimal control. Finally, under the convexity assumption on the Hamiltonian function, we show that this optimality condition is also sufficient.

Keywords: Control theories, Intelligent control
Abstract: Model predictive control (MPC) is attractive in engineering for its explicit constraint handling. To deal with ubiquitous disturbances, many strategies are utilized to improve the disturbance-rejection performance. However, they cannot be effectively incorporated into MPC, causing deviation from the optimal solution. To address this issue, this paper proposes an incremental MPC method based on an equivalent input disturbance (EID) predictor. First, the closed-loop system is reorganized into an estimation-error subsystem and a controloriented subsystem, and the control-oriented subsystem is further decomposed into tracking and disturbance-rejection components. Further, both subsystems are formulated and integrated into a unified MPC framework. Moreover, the stability of the closedloop system is analyzed. Finally, a case study on a rotary motor system and comparisons with other methods validate the effectiveness and superiority of the proposed method.

Keywords: Nonlinear control and applications
Abstract: Traditional local linearization-based control methods cannot capture global nonlinear dynamics. Although the Koopman operator enables globally linear models, the resulting models are typically high-dimensional and contain unstable modes, which cannot be directly controlled. To address this issue, a robust control method that jointly enforces stability constraints and incorporates Equivalent Input Disturbance (EID) is proposed. First, the finite-dimensional approximation of the Koopman operator is constructed using observable functions identified via the extended dynamic mode decomposition (EDMD). Second, the asymptotic stability of the unobservable subsystem is guaranteed by minimizing its matrix's Frobenius norm to place all eigenvalues strictly inside the unit circle. Subsequently, balanced truncation is employed to obtain a low-dimensional, controllable, and observable reduced-order model. Modeling errors from finite-dimensional approximation and reduction are treated as equivalent input disturbances and actively rejected via an EID-based feedforward compensator. Finally, simulation results on the Van der Pol oscillator verify that the proposed approach eliminates steady-state tracking error caused by model approximation, demonstrating effective and superior tracking performance for nonlinear systems.

Keywords: Control theories
Abstract: Aiming at the tracking control of linear systems under unknown disturbances, conventional methods have difficulty simultaneously guaranteeing prescribed error constraints and disturbance rejection under large transient disturbances. In particular, when disturbances exceed the compensation capability of the Equivalent Input Disturbance (EID) observer, fixed-boundary Prescribed Performance Control (PPC) schemes may suffer from boundary violation or unstable. To address this issue, a Prescribed Performance Control with Event-Triggered Dynamic Boundary Adjustment based on EID (PPC-ETM-EID) is proposed. Different from conventional PPC methods with fixed or continuously adaptive boundaries, the proposed approach reconstructs the performance boundary only when the tracking error approaches the constraint limit. By coordinating disturbance estimation, boundary evolution, and controller update, the method prevents performance violation while reducing unnecessary conservatism. Meanwhile, the EID observer estimates unknown disturbances online, and a proportional–derivative control law is designed based on transformed. Simulation results demonstrate that the proposed method effectively maintains tracking errors within prescribed bounds, improves disturbance.

Keywords: Control theories
Abstract: This article presents a dissipative problem of linear uncertain time-delay systems. First, the system control structure is constructed using a modified equivalent-input-disturbance (MEID) approach. The MEID method is applied to obtain disturbance estimate of the system, and design the disturbance rejection control rate. Then, an appropriate Lyapunov-Krasovskii functional is considered. The stability conditions of the system are given in the form of Linear-Matrix-Inequality (LMI). And the dissipative performance for the closed-loop system is satisfied. Building on this foundation, an output feedback controller is explained. Ultimately, the benefits of the proposed methodology are demonstrated through a numerical simulation.

Keywords: Deep learning and machine learning, Artificial intelligence
Abstract: Electric vehicles (EVs) charging's stochasticity aggravates grid fluctuations, while new stations face cold-start challenges and existing models lack interpretability. Therefore, this paper proposes Gated Adaptive Temporal-Elastic Kolmogorov-Arnold Network (GATE-KAN), a high-accuracy interpretable forecasting method for cold-start scenarios. It uses a frozen temporal foundation model as the base predictor, with a dynamically gated KAN residual adapter for lightweight bias correction. The adapter integrates separable guided pruning and economic monotonicity constraints, enabling analytical derivation of real-time price elasticities. A multi-criteria donor migration strategy and fading adaptive Kalman filter are designed for cold-start and online calibration. Experiments on ST-EVCDP show GATE-KAN reduces RMSE by 7.39% over the suboptimal fine-tuned model, with better generalization and convergence.

Keywords: Control theories, Artificial intelligence, System identification and modelling
Abstract: Hydrodynamic disturbances severely degrade the torque tracking accuracy of conventional exoskeleton controllers during underwater locomotion. For effective assistance of human-robot interaction, this paper proposes a hierarchical control method for a knee-joint exoskeleton. First, an adaptive oscillator is used to extract swimmer phase from inertial measurements to generate a nominal synchronized trajectory. Then an admittance controller is incorporated to convert this trajectory into a compliant desired trajectory for human-robot interaction. To achieve accurate trajectory tracking despite fluid drag and added mass uncertainties, a composite controller by combining a sliding mode controller (SMC), a sliding mode observer (SMO), and a radial basis function (RBF) neural network is designed. Further, the semi-global uniformly ultimately bounded (SGUUB) stability of the developed closed-loop system is analyzed. Finally, simulation results show under 20% model perturbations, the proposed method reduces the trajectory-tracking root-mean-square error (RMSE) from 0.2665 rad to 0.0362 rad, an 86.4% improvement over a baseline sliding mode controller.

Keywords: Control theories, Motion and vibration control, System identification and modelling
Abstract: Estimations of the maximum responses of a high-rise active base-isolated building subjected to wind force rely on time-history simulation, which is complex and time-consuming. This study proposes a simple estimation method for estimating the maximum responses and maximum control force for a high-rise active base-isolated building on across-wind direction. Then, the scope of application range of the proposed method is discussed. In this study, we established some control parameters and verified whether the estimation method has a scope of application. As a result, when the displacement of base-isolated story is evaluated, or the displacement and the velocity of all stories are evaluated, we confirm that the estimation accuracy of the proposed method is good.

Keywords: Nonlinear control and applications, Control theories, System identification and modelling
Abstract: This paper proposes a tube-based model predictive control (MPC) framework for constrained nonlinear Lur’e systems using integral quadratic constraints (IQCs). In contrast to existing tube MPC approaches based on Lipschitz assumptions or bounded disturbances, the proposed method exploits dissipativity theory and IQCs to synthesize a robust stabilizing feedback law for the error dynamics. The error system is modeled as a nonlinear feedback interconnection satisfying prescribed IQCs, which leads to linear matrix inequality conditions guaranteeing absolute stability with respect to slope-restricted nonlinearities. Based on these conditions, an effective disturbance description is derived and used to construct robust positively invariant tubes, ensuring recursive feasibility and robust constraint satisfaction.

Keywords: Control theories, Industrial applications
Abstract: This paper establishes a unified framework for associating the proportional-integral-derivative (PID) controller, input/output state augmentation, controllability, internal model principle (IMP), and Rosenbrock system matrix. It is unveiled that output-integrator-based formulations (e.g., classical (position form) PID and servo-linear quadratic regulator (LQR) impose a structural constraint equal to the absence of invariant zeros at the origin, whereas input-integrator formulations (velocity-form PID) do not. A precise relationship is established between coordinate transformations, increased controllability, and the potential of steady-state tracking. The findings reveal fundamental constraints of PID controllers and explain why input-augmented solutions have more structural robustness in state space representation.

Keywords: Industrial applications, Artificial intelligence, Measurement and instrumentation
Abstract: Predictive condition monitoring of critical process parameters is essential for ensuring product quality and stable equipment operation in cold rolling processes. To overcome this challenge, this study develops a hybrid deep learning model, namely DALA-MSeq2Seq. First, a multi-scale gated recurrent unit-based sequence-to-sequence model is designed, which employs parallel 1D convolutional layers to extract multi-scale temporal features from input sequences and perform multi-step prediction. Subsequently, to achieve automatic optimization of model hyperparameters, the differential artificial lemming algorithm is proposed, which incorporates a dynamic differential evolution strategy with stagnation counter to avoid falling into local optima. Experimental evaluation on a real cold rolling production dataset demonstrates that compared to six baseline models, the proposed method exhibits the most promising performance with MSE=0.0006, R2=0.9256, MAE=0.0143, and RMSE=0.0253 in 6-step thickness deviation prediction.

Keywords: Intelligent control, Industrial applications, Deep learning and machine learning
Abstract: In industrial processes, stable control of key performance indicators (KPIs) is crucial for optimizing raw material utilization and ensuring product quality. However, the presence of multiple operating modes poses significant challenges to process control. To address this issue, a knowledge-guided soft actor–critic (SAC) method for industrial process control under multiple operating modes, termed OKG-SAC, is proposed. Within the SAC framework, process expert knowledge is incorporated to guide the agent toward more mode-specific action ranges under different operating modes, while balancing policy convergence performance and optimality. A simulation case study based on a practical zinc leaching process (ZLP) is conducted to validate the effectiveness of the proposed method.

Keywords: Control theories, Artificial intelligence, Industrial applications
Abstract: This paper proposes an output-feedback based adaptive reinforcement learning (RL) backstepping control structure to achieve stable and precise control of a permanent magnet synchronous motor (PMSM) under model uncertainty and disturbance conditions. The proposed method combines parameter identification using an improved Levenberg–Marquardt (iLM) algorithm with an extended state observer (ESO) to estimate unmeasured states and disturbances in real time and incorporate them into the controller. This enables improved model accuracy and disturbance compensation without requiring full state measurement. Furthermore, Lyapunov-based stability analysis theoretically guarantees semi-globally ultimate uniformly boundedness (UUB) by the input-to-state stability (ISS) property for the closed-loop system. Experimental results on an actual PMSM demonstrate that the proposed structure provides improved tracking performance and disturbance rejection capability compared to existing methods, confirming that the parameter identification and adaptive learning processes converge stably. These results suggest that the proposed control structure offers practical and robust performance in PMSM application environments.

Keywords: Measurement and instrumentation, System identification and modelling, Control devices, sensors and actuators
Abstract: The Interacting Multiple Model (IMM) filter is a Bayesian tracking algorithm, effective for targets that frequently switch between different motion modes. However, its performance critically depends on the internal transition probability matrix (TPM), which in conventional approaches must be obtained from prior information or trained offline. To overcome this limitation, we introduce IMM OAGD (IMM with Online Adaptive Gradient Descent) algorithm that dynamically optimizes the TPM during the filter’s recursive operation. Firstly, we design the algorithm framework based on IMM OAGD. Secondly, we propose a novel weighted loss function that solely relies on the observation sequence to update the parameters. This requires neither ground-truth system state nor pre-collected datasets, and only assumes that the system dynamic and measurement models are differentiable. Finally, simulation results demonstrate the effectiveness and robustness of IMM OAGD, particularly in tracking targets with complex maneuvering patterns.

Keywords: Autonomous vehicles, Control theories, Cyber-physical systems and security
Abstract: This paper studies the problem of safe control of unmanned aerial vehicle (UAV) formation in environments with unforeseen dynamic obstacles. Inspired by the water drop model, an improved artificial potential field (APF) method is proposed. To further enhance the responsiveness of UAVs to unforeseen dynamic obstacles, the obstacle's velocity and volume are introduced to optimize the potential field. Additionally, leveraging the interaction of water waves, an auxiliary potential field is incorporated to address the local minima problem inherent in traditional APF methods. Finally, numerical simulations validate the safety and effectiveness of the proposed method in enabling UAVs to handle unforeseen dynamic obstacles.

Keywords: Control theories, System identification and modelling
Abstract: A DeePC design method based on source selection and affine output correction is proposed for a group of devices that are homogeneous but not identical. Instead of constructing DeePC only from the short data of the target device, the proposed method exploits long data from an existing device whose response is close to that of the target. The source device is selected on the basis of the one-step-ahead prediction RMSE computed on the short data of the target device. The basic design directly uses the long data of the selected device. In the extended design, the long output sequence is first subjected to affine output correction. The long-data-based designs are also allowed to adopt a larger past horizon so that the benefit of the data length is reflected in the controller design. The effectiveness of the proposed method is verified by leave-one-device-out numerical simulations for a homogeneous device group consisting of ten stable second-order SISO continuous-time LTI systems. Each plant is sampled under zero-order-hold inputs, DeePC is applied to the resulting input-output data, and the closed-loop response is evaluated in continuous time including intersample behavior. In the numerical results, the tracking IAE of the short-data-only design deviates markedly from those of the three designs using long data for some targets. In the largest-gap case, its IAE exceeds that of the better of the two practical long-data-based designs by 2.10 and is 26.6% larger than the mean IAE of

Keywords: Control theories, Cyber-physical systems and security, Robotics and swarm intelligence
Abstract: This paper investigates the consensus problem of multi-agent systems (MASs) whose communication topologies are subject to false-data-injection (FDI) attacks and proposes a distributed control scheme based on discrete-time state observers. Because the attacks distort the information flow between the followers and the leader over the communication links, we first design a distributed observer that estimates the leader’s state and the local tracking errors by using the relative outputs of neighboring agents. Building on this result, we further develop an event-triggered control protocol for discrete-time MASs that enables the followers to track the leader’s state trajectory effectively even under FDI attacks. The protocol reduces the triggering frequency and saves communication resources while guaranteeing mean-square consensus of the discrete-time MASs. Finally, numerical simulations are provided to demonstrate the effectiveness of the proposed approach.

Keywords: Control theories, Mechatronics, Robotics and swarm intelligence
Abstract: This study presents an exact discretization method for fractional-order systems with inputs based on the augmented system approach. A formulation of fractional-order dynamics using the Caputo derivative is presented, together with the Mittag--Leffler function, which is essential for representing exact solutions. Based on this formulation, an augmented state representation is introduced to express the continuous-time response under external inputs in a structured manner, and a discrete-time update equation is derived without relying on the semigroup property, which does not hold for Mittag--Leffler functions, thereby overcoming a fundamental limitation in conventional discretization approaches. The proposed method provides a systematic approach for computing the trajectories of fractional-order systems in both continuous and discrete time domains, while preserving key characteristics such as memory effects. As a result, it offers a practical and efficient basis for developing digital control and simulation techniques for fractional-order systems.

Keywords: Control theories, Motion and vibration control, Autonomous vehicles
Abstract: We consider a problem of asymptotic stabilization of a linear system state within a given stationary ellipsoid with simultaneous evasion of interception by an arbitrary number of moving ellipsoids. A nonlinear control law in closed form that locally solves such task is derived using quadratic programming (QP) framework with control Lyapunov function (CLF) and control barrier function (CBF) constraints. Conservative domain of initial conditions, from which the problem under consideration can be solved using the proposed control law, is obtained. Theoretical result is illustrated by a numerical example, in which a linear system counteracts interceptors based on parallel navigation method.

Keywords: Control theories, Nonlinear control and applications
Abstract: In this paper, we propose a modified saddle-point dynamics to solve inequality-constrained distributed optimization problems on strongly-connected general digraphs. To overcome the limitations of a weight-unbalanced digraph, we propose a novel continuous-time weight-balancing dynamics, with each iterate of weights being simultaneously used to solve the rest of the saddle-point dynamics consisting of the projection-based non-smooth nature for inequality constraints. To analyze the stability analysis of the proposed dynamical system, we first show the trajectory of the weight-balancing dynamics reaches steady-state weights that balance the digraph exponentially. We next show that the optimal solution for the general digraphs are the Lyapunov stable equilibrium points of the proposed dynamics and finally, every trajectory of the proposed dynamics asymptotically reaches to an equilibrium point.

Keywords: Control theories, Intelligent control, System identification and modelling
Abstract: This paper presents a deep learning-based finite-dimensional nonlinear model predictive control (NMPC) strategy for a class of ordinary differential equation (ODE)-Euler-Bernoulli beam (EBB) cascaded systems. To address the challenges posed by the infinite-dimensional nature of such systems, a deep autoencoder is first employed to extract latent representations of the EBB subsystem through dimensionality reduction. Subsequently, the sparse identification of nonlinear dynamics algorithm is utilized to identify the reduced-order model (ROM). By integrating the identified ROM with the ODE subsystem, a unified coupled model and its corresponding error dynamics are formulated. This error system then serves as the internal prediction model for the NMPC framework, enabling asymptotic state tracking and vibration suppression. Finally, numerical studies on a flexible morphing hypersonic vehicle demonstrate effectiveness of the proposed scheme.

Keywords: Control theories, Nonlinear control and applications
Abstract: This paper proposes a self-triggered distributed optimization algorithm based on the Zero-Gradient-Sum (ZGS) framework for continuous-time multi-agent systems. While distributed optimization is critical for coordinating large-scale networks, traditional event-triggered mechanisms suffer from continuous channel monitoring requirements that significantly drain onboard energy and communication resources. To fundamentally resolve this bottleneck, our proposed self-triggered mechanism enables each agent to proactively compute its subsequent communication instant using exclusively local information, allowing transceivers to remain in a low-power state between triggers. Rigorous theoretical analysis demonstrates that the algorithm drives the system to the exact global optimum with an exponential convergence rate. Furthermore, the existence of a strictly positive lower bound for inter-execution times is mathematically proven, thoroughly excluding Zeno behavior. Numerical simulations, contextualized within an abstracted satellite network model, verify that the proposed algorithm secures absolute optimization accuracy while maintaining an energy-efficient sparse communication pattern.

Keywords: Control theories, Nonlinear control and applications, Artificial intelligence
Abstract: This paper presents an adaptive impedance control method to address the safety issues of electrically driven flexiblejoint robotic manipulators (EFRMs) during interactive tasks, and combines it with finite-time fuzzy adaptive control to achieve dualtrack tracking control of trajectory and contact force. Impedance control in contact space is considered, and new adaptive laws are designed to estimate unknown environmental information. A finite-time anti-saturation signal is adopted to prevent excessive saturation of the actuator input. A nonlinear state-independent function is applied to ensure that the output position of the endeffector does not violate the asymmetric time-varying constraints. Based on Lyapunov stability theory, it is rigorously proved that the designed controller guarantees uniform boundedness of all closed-loop system signals within a finite-time. The effectiveness of the proposed scheme is verified through the simulation results.

Keywords: Control theories, Control devices, sensors and actuators, Cyber-physical systems and security
Abstract: For a class of LTI systems driven by multiple state-dependent perturbations, we provide synthesis tools to design optimal monitoring schemes. We derive tools to construct system monitors that account for system and output perturbations and provide optimal monitoring rules in terms of minimum volume residual-based ellipsoids. That is, if the signals to be monitored lead to the residuals leaving these ellipsoids, alarms are raised, indicating potential faults. We derive outer ellipsoidal bounds for the set of reachable trajectories that peak-bounded perturbations can induce, and project them onto the residual hyperplane. We design the dynamics of the monitor so that this projection contains, as tightly as possible, the residual trajectories that standard perturbations (i.e., no faulty signals) can induce in the system.

Keywords: Autonomous vehicles, Artificial intelligence, Deep learning and machine learning
Abstract: Road potholes remain a persistent road-safety problem because damaged road surfaces reduce driving comfort, increase vehicle maintenance costs, and may threaten traffic safety. This paper proposes a simulation-based late-fusion framework for pothole detection and severity estimation using an RGB camera, LiDAR, and IMU in CARLA. A custom 3D pothole environment is developed in Unreal Engine 4, while a YOLOv11 detector trained on pothole images collected from different cities in Indonesia is used to localize pothole candidates in RGB frames. LiDAR is then used for region-of-interest (ROI)-guided depth estimation and severity classification based on local road geometry, while the IMU provides post-traversal physical validation through shock- and jerk-based responses. Unlike vision-only approaches, the proposed framework separates visual localization, geometric verification, and inertial validation at the decision level. The results show that the RGB module can generate candidate ROIs for the fusion pipeline, the LiDAR module captures the relative trend of pothole depth and rejects visually detected false positives without valid geometric depression, and the IMU module provides supporting physical validation during traversal. The final system produces actionable outputs in the form of textit{SAFE}, textit{SLOW DOWN}, and textit{CHANGE ROADS}.

Keywords: Autonomous vehicles
Abstract: This paper proposes the Spatial Risk Evaluation Method (SPREME), which uses Impact Vertex Projection (IVP) to safely execute an autonomous lane change. The proposed method is a projection-based regional threat evaluation, and it simplifies a two-dimensional crash situation to a one-dimensional line. We predict the likelihood of a crash and the impact point by considering the combined relative position, velocity, and orientation of objects. The validity was confirmed by comparing the proposed method with an existing one, and we observed that it outperforms the current approach with faster judgment. In addition, our scheme can determine the crisis even in a threatening situation, which the current strategy cannot respond to.

Keywords: Autonomous vehicles, Artificial intelligence, Control devices, sensors and actuators
Abstract: Driven by the expanding low-altitude economy, autonomous ground handling is pivotal for streamlining air-ground logistics. This paper presents a robust Visual-Language Navigation framework for omnidirectional ground vehicles tasked with UAV transport. Addressing the challenges of semantic ambiguity and dynamic environments, we propose a hierarchical architecture. First, we introduce a two-stage Large Language Model parsing strategy that decouples semantic simplification from structured formalization, significantly mitigating hallucination risks. Second, we implement a multi-modal perception module that grounds 2D visual detection into 3D LiDAR point clouds via density-based clustering for precise landmark localization. Third, a spatial-temporal trajectory planner utilizing MINCO parameterization within Safe Flight Corridors (SFC) is employed, augmented by virtual force fields to ensure kinematic feasibility and dynamic safety. Finally, the approach was validated in the 3rd Meituan Low-Altitude Economy Intelligent Flight Management Challenge, where it demonstrated exceptional sim-to-real robustness and secured the Runner-up position (2nd Place) with superior Success Weighted by path length performance.

Keywords: Autonomous vehicles
Abstract: The deployment of autonomous and remote control methodologies on research vessels requires rigorous pre-deployment validation to ensure safety and reliability in the complex marine environment. This paper presents a comprehensive framework comprising: (i) definition of technical limits and system interfaces; (ii) high-fidelity simulation using three-degree-of-freedom ship dynamics, propeller hydrodynamics, extended Kalman filter navigation, and PID control; (iii) A system-level fault detection, isolation, and recovery (FDIR) architecture based on normalised innovation squared (NIS) monitoring. Simulation results for a 20% and 40% port propeller thrust loss scenarios (propeller damaged) demonstrate fault detection. The framework provides a principled, safety-critical pathway from laboratory validation to future controlled field deployment of autonomous vessel systems.

Keywords: Autonomous vehicles, Control theories, Control devices, sensors and actuators
Abstract: Vehicle reliability and safety are of fundamental importance to the development and deployment of autonomous driving technologies. This paper proposes a safety-critical control method considering adaptive output constraints and actuator faults for the safety control problem of autonomous vehicles. First, considering actuator faults, system uncertainties, and external disturbances, a vehicle dynamics model is established.. Then, a composite estimation module consisting of a nonlinear disturbance observer(NDO) and an radial basis function neural networks(RBFNNs) is constructed to jointly compensate for the effects of system uncertainties, external disturbances and actuator faults. Furthermore, by comprehensively considering external environmental factors and vehicle operating conditions, a state-dependent barrier Lyapunov function(BLF) is developed, and a distance-triggering mechanism coupled with vehicle states is introduced, thereby achieving adaptively adjustable dynamic output constraints and enabling automatic switching between constrained and unconstrained area. The BLF is further embedded into a backstepping control framework to ensure strict satisfaction of the dynamic output constraints. In addition, the semiglobal bounded stability of the system is established via the direct Lyapunov method. Finally, the simulation experiments are conducted to verify the effectiveness of the proposed safety control scheme.

Keywords: Nonlinear control and applications, Autonomous vehicles, Control theories
Abstract: This paper proposes a reinforcement learning (RL) –enhanced model predictive control (MPC) framework for longitudinal control of autonomous vehicles. The proposed approach integrates an actor–critic RL module to adapt MPC cost parameters online, thereby improving performance across varying driving conditions while preserving MPC's constraint-handling and stability properties. By restricting learning to cost adaptation, the control structure and safety guarantees of the underlying MPC formulation are maintained. The resulting RL-MPC framework provides a practical and computationally efficient solution for performance-adaptive longitudinal control in autonomous driving.

Keywords: Autonomous vehicles, Control theories, Intelligent control
Abstract: Autonomous underwater gliders are increasingly deployed for versatile and persistent ocean monitoring appli- cations. Since these vehicles rely on limited onboard battery capacity, minimizing energy consumption is critical to extending their exploratory endurance. While traditional controllers are typically time-triggered, event-triggered controls provide better actuator usage management. However, current methods rely solely on spatial tracking errors and ignore the vehicle’s internal energy state. This paper proposes a novel energy-aware event- triggered geometric proportional-derivative control framework. Formulated directly on the Special Euclidean group SE(3), the geometric PD controller ensures robust and computationally efficient tracking. The proposed design explicitly incorporates the battery’s state of charge into the triggering condition to dynamically adjust the control update frequency. The framework is evaluated in simulation for the pitch and depth tracking of a Seawing underwater glider. Comparative results demonstrate that the proposed energy-aware ET-Geometric PD controller achieves more than 45% energy savings and decreases control updates by 31% compared to a conventional time-triggered geometric PD controller over the same mission duration.

Keywords: Autonomous vehicles, Deep learning and machine learning, Adaptive systems
Abstract: This paper presents a comprehensive investigation of transfer learning techniques for Model Predictive Control(MPC) cloning across different friction coefficients in autonomous vehicle applications. We implement and compare three state-of-the-art transfer learning approaches: CORrelation ALignment (CORAL), Continue Training, and Ensemble Learning. Our experiments demonstrate that transfer learning can achieve up to 27.4% improvement in prediction accuracy with only 500 target domain samples. The Continue Training method emerges as the most effective approach, successfully transferring knowledge from high-friction (μ = 0.7) to low-friction (μ = 0.3) conditions. This work contributes to the field by providing empirical evidence of transfer learning efficacy in safety-critical automotive control systems and establishing best practices for domain adaptation in MPC imitation learning.

Keywords: Artificial intelligence, Control theories
Abstract: This manuscript presents a reinforcement learning (RL) based approach for fine-tuning proportional–integral (PI)controllers in decoupled multi-input multi-output (MIMO) process control systems. Conventional PI tuning methods typically rely on explicit plant models or approximations of process dynamics, which can limit their applicability to complex or uncertain systems. In contrast, the proposed method adopts a model-free framework, where an RL agent incrementally adjusts controller gains by interacting with the closed-loop system. The control action for each loop is formulated by augmenting the existing PI controller structure with additional gain corrections from the RL agent, enabling performance enhancement without altering the underlying control architecture. The method was validated on several benchmark MIMO processes, including the VL column, Wood & Berry, Wood & Wardel and ISP plant. Simulation results demonstrate that the RL-tuned controllers achieve faster transients, reduced integral error indices (IAE and ISE), and enhanced robustness under plant–model mismatch and external disturbances compared to existing tuning strategies.

Keywords: Artificial intelligence, Deep learning and machine learning
Abstract: Domain-incremental learning (Domain-IL) aims to adapt models to sequential data domains with distribution shifts while mitigating catastrophic forgetting of prior knowledge, particularly when domain labels are unavailable during inference. A fundamental challenge in this paradigm is the inter-domain interference, where domain-specific features often act as confounders that degrade generalization. In this paper, we propose a novel framework, Causal Knowledge Distillation for Incremental Learning (CKD-IL), specifically evaluated on the BONGARD-LOGO benchmark. By formulating a Structural Causal Model (SCM), our approach identifies and decouples domain-specific confounders from invariant representations. We then employ a dual-stage knowledge distillation mechanism to preserve both universal and domain-specialized knowledge. Experimental results demonstrate that CKD-IL achieves state-of-the-art (SOTA) performance, effectively mitigating interference and enabling robust generalization to complex abstract concepts in a sequential learning setting.

Keywords: Artificial intelligence, Deep learning and machine learning, Robotics and swarm intelligence
Abstract: Image restoration focuses on reconstructing high-quality images from degraded inputs affected by distortions such as noise, rain, or haze. Although recent deep learning-based methods have shown strong restoration capabilities, many existing approaches are still designed for particular degradation categories or depend on prior information about the type of corruption. To overcome these limitations, we introduce a unified all-in-one image restoration framework built upon a convolutional mixture-of-experts (C-MoE) architecture. The proposed model dynamically selects specialized convolutional experts to process different degradation characteristics. In particular, we introduce a convolutional MoE that incorporates both point-wise and depth-wise expert routing, enabling the network to effectively handle pixel-wise and spatial feature variations. Comprehensive evaluations on denoising, deraining, and dehazing benchmarks show that the proposed approach delivers competitive restoration performance with efficient computational cost. The results show that our approach effectively generalizes across multiple restoration tasks and improves robustness under diverse image degradation conditions.

Keywords: Control devices, sensors and actuators, Deep learning and machine learning, Artificial intelligence
Abstract: Wildlife intrusion in plantation environments can cause severe crop damage and significant economic losses, especially in plantations located near forest areas. This article proposes an AIoT-based smart repellent system that uses virtual fences defined by cameras for automatic detection and mitigation of wildlife intrusion without relying on physical barriers. First, infrared cameras are installed along the plantation boundaries to continuously record image sequences of the monitored area. These images are processed in real-time using an enhanced YOLOv12-based detection framework, called YOLO-WD (wildlife detection), which improves feature extraction and multi-scale feature fusion to reliably detect wildlife under complex backgrounds and varying lighting conditions. When wildlife is detected entering a predetermined virtual fence area, a sound deterrent mechanism is automatically activated, using diverse sound patterns to prevent further intrusion. The proposed system is implemented on an edge computing platform using jetson orin nano and integrated with real-time monitoring and mobile notification services via LINE chatbot to ensure timely user awareness. Experimental results show that the proposed YOLO-WD model achieves a F1 score of 98.27% and a mAP@0.5 of 97.31% while maintaining real-time inference performance. Further field application results confirm the effectiveness and practicality of the proposed system for real-world orchard protection.

Keywords: Deep learning and machine learning, Artificial intelligence
Abstract: This paper proposes Vmamba-DETR, a novel object detection network built upon RF-DETR (Real-Time Feature-Enhanced DEtection TRansformer) and Vision Mamba (Vmamba). The proposed method aims to improve the detection of maritime targets such as vessels and buoys in complex ocean environments. The network integrates architectural designs from Deformable DETR and Mamba. Vmamba is adopted as the backbone, and its core component, the 2D Selective Scan (SS2D) Module, is incorporated to improve computational efficiency. In addition, a Gate-Controlled Vision Mamba Mixer is introduced to enable more effective fusion of spatial and sequential information for vision-based detection tasks. An auxiliary positional supervision loss is further employed to guide spatial feature learning and enhance detection accuracy. According to network scale, two variants are presented: Vmamba-DETR-Large (Vmamba-DETR-L) and Vmamba-DETR-Base (Vmamba-DETRB). Benefiting from the Transformer-based detection framework, the proposed approach improves global context modeling and robustness to partial occlusion. On a custom marine dataset, Vmamba-DETR-L achieves an mAP50 of 0.978, outperforming multiple YOLOv12 variants and Faster R-CNN. Meanwhile, Vmamba-DETR-B demonstrates superior performance over RFDETR, Faster R-CNN, and YOLOv12 with comparable parameter scales. On the public Buoy-Onboarding dataset, experimental results illustrate that Vmamba-DETR achieves competitive or state-of-the-art mAP50 and mAP50:95.

Keywords: Deep learning and machine learning, Artificial intelligence, Computational intelligence
Abstract: Multi-modal emotion recognition aims to perceive and interpret human affects by integrating heterogeneous modalities, including textual, visual and acoustic signals. However, multi-modal data are often incomplete due to sensor failures or transmission losses in real-world scenarios, which substantially degrades the performance of conventional models. In this paper, we propose a novel Transformer-based conditional diffusion network (TCDNet) for incomplete multi-modal learning in conversations. TCDNet explicitly introduces modality-specific information and cross-modal conditions into the diffusion denoising process. This design enables robust modality completion under missing-modality settings and facilitates discriminative representation learning. Compared with prior work, TCDNet further aligns cross-modal semantics with modality-consistent representations, facilitating reliable feature recovery during denoising. We evaluate TCDNet on two benchmark conversational datasets. Experimental results demonstrate that TCDNet consistently outperforms state-of-the-art methods for incomplete multi-modal learning.

Keywords: Deep learning and machine learning, Artificial intelligence, Computational intelligence
Abstract: This study proposes an integrated framework for segmenting and classifying rice status from UAV RGB images at the near-harvest stage. In the first stage, Segment Anything Model 2 (SAM2) was used to extract object regions from the original images. Several key parameters were selected for segmentation, including points_per_side = 24, pred_iou_thresh = 0.7, stability_score_thresh = 0.8, and min_mask_region_area = 3000. The valid segmented regions were then used to build a dataset with three classes: Green, Lodged, and Ripened. In the second stage, the DeiT-Small model was applied to classify each segmented image region. Experimental results showed that the model learned stably and converged after about 17 epochs. The proposed framework achieved an overall classification accuracy of 98.62%, with F1-scores of 0.9861, 0.9862, and 0.9864 for the Green, Lodged, and Ripened classes, respectively. The results also showed that the combination of SAM2 and DeiT-Small can support accurate classification in small local regions of rice-field images. However, some limitations were still observed, especially in lodged areas with irregular shapes and in ripened areas that were more fragmented or easily confused near transition zones. Overall, the proposed framework shows strong potential for rice-status analysis and pre-harvest rice field assessment using UAV imagery combined with modern deep learning models.

Keywords: Deep learning and machine learning, Intelligent control, Industrial applications
Abstract: Feedforward–feedback control (FFC–FBC) architectures are widely adopted for setpoint tracking and disturbance rejection due to their structural simplicity and robustness. Nevertheless, the performance of conventional linear FFC–FBC designs deteriorates when applied to nonlinear chemical processes operating over wide ranges. Existing data-driven controller tuning methods offer limited theoretical guidance, restricting their ability to generalize across varying operating conditions. This work develops an autonomous, control-informed reinforcement learning (RL) framework for systematic optimization of FFC–FBC parameters using a divide–conquer–integrate paradigm. The divide stage decomposes the nonlinear plant into locally linearized operating regions. In the conquer stage, region-specific RL agents are trained to optimize controller parameters, with learning shaped by control-theoretic constraints and performance objectives inherent to FFC–FBC structures, thereby improving learning efficiency and convergence properties. The integrate stage synthesizes the local policies into a unified global controller, enabling consistent performance over the full nonlinear operating domain. Theoretical insights into the learning structure and integration mechanism are provided, and the proposed framework is validated through an industrial case study, demonstrating robustness and generalization across multiple operating regions.

Keywords: Robotics and swarm intelligence, Deep learning and machine learning, Measurement and instrumentation
Abstract: Ensuring stable visual localization for Unmanned Aerial Vehicles (UAVs) is critical in GNSS-denied environments. Although existing retrieval-based methods provide a foundation, they often rely on idealized datasets where queries are perfectly centered over satellite patches. To address this, we propose a robust spatially aware framework and a new dataset incorporating random spatial offsets for real-world deployment where perfect alignment is impossible. Our approach integrates an atmospheric simulation pipeline to enhance feature invariance against environmental and hardware noise. To bridge the gap between discrete retrieval and continuous geographic space, we introduce Spatial Label Smoothing and Spatial Soft Triplet Loss. These components incorporate geographic distances directly into training, allowing the loss function to avoid penalizing the model for landscapes of similar appearance that are geographically distant. This prevents the network from creating unnatural embedding patterns to distinguish identical terrain. The experimental results of our dataset show improvements over baseline of 3.38% in R@1, 1.16% in R@5, 66.6% in MAE and 2.87% in Median error. Robustness tests using historical satellite imagery further demonstrate gains of 10.6% in R@1 and 70.2% in Median error. Our approach shows that integrating spatial constraints significantly reduces large-scale positioning errors, providing a reliable anchor for navigation in GNSS-denied flight.

Keywords: Robotics and swarm intelligence, Computational intelligence, Intelligent control
Abstract: This paper proposes an evolutionary control approach for implementing navigation of a real hexapod robot in unknown environments. The control approach consists of lower-level and higher-level controllers. In the lower-level control, a fully connected recurrent neural network (FCRNN) is responsible for generating the basic walking gait of the robot. The FCNNN sends six signals to control the six hip joints of the robot with the two control objectives of fast walking speed and straightness in forward walking. In the higher-level control, two fuzzy controllers (FCs) are employed. One FC controls the robot to execute the obstacle boundary-following (OBF) behavior with the objectives of fast walking speed and maintaining a proper distance to an obstacle. The other FC controls the robot to execute the target searching (TS) behavior with the objectives of fast walking speed and low angle deviations between the robot and the target. Based on the sim-to-real policy, the above three controllers are optimized using a multiobjective evolutionary computation algorithm through simulations. Finally, the FCRNN and FCs learned from simulations are applied to navigate a real robot in an unknown environment.

Keywords: System identification and modelling, Robotics and swarm intelligence, Motion and vibration control
Abstract: Physical parameters in robot dynamics model have to be concerned in order to design and develop model-based control algorithm. This work present the novel parameters identification technique used by five degree of freedom serial manipulator in dual-arm robot system. An evolutionary algorithm called particle swarm optimization and Haar wavelet are combined together to estimate the physical parameters in the dynamic model, this problem are considered as nonlinear optimization problem. Particle swarm optimization algorithm will be used to search for the best-fit parameters. Computation cost and effect of high frequency disturbance are reduced by representing all state vectors in dynamics model by discrete Haar wavelet.

Keywords: Control theories, Intelligent control
Abstract: This study proposes a new robust T-S fuzzy model for networked control systems. Based on the presented robust T-S fuzzy model, an LMI-based networked controller for a mobile robot is developed. Furthermore, novel robust stability conditions are derived in terms of linear matrix inequalities (LMIs) using the Lyapunov–Krasovskii theorem. Compared with existing T-S fuzzy networked control systems, the number of fuzzy rules is reduced from six to two, thereby decreasing the computational time required by the microchip. Moreover, the number of LMI-based stability conditions is reduced, which enlarges the feasible solution space for the control gains. Both simulation and experimental results show that the proposed novel T-S fuzzy networked controller achieves superior robustness performance compared with existing approaches.

Keywords: Brain-computer interfaces, Adaptive systems, Neuro and bioinformatics
Abstract: Wireless electroencephalography (EEG) has developed rapidly in recent years, reflecting a transition from conventional laboratory environments toward portable, wearable, and long-duration monitoring scenarios. For real- world attention monitoring, such systems must be not only portable and easy to use, but also capable of providing physiologically interpretable EEG measurements. To address this need, a portable forehead EEG system was developed for real-time adaptive attention monitoring in real-world settings. The proposed system is built on NEEG-2, a self-developed two-channel forehead EEG headband integrating gel-free graphene dry electrodes and a custom printed circuit board (PCB) for wearable EEG acquisition. To verify its practicality, NEEG-2 was validated against an FDA-cleared EEG device under resting, blinking, and teeth-grinding conditions. Experimental results showed strong inter-device agreement in both the time and frequency domains, together with consistent EO–EC alpha-band modulation, supporting the physiological validity of the proposed system. Based on an Attention Network Test (ANT)-inspired paradigm, a real-time adaptive attention monitoring framework was further established using frontal EEG activity for individualized calibration and online attention estimation. These results demonstrate the potential of the proposed system as a practical and interpretable forehead EEG platform for wearable attention monitoring beyond conventional laboratory settings.

Keywords: Autonomous vehicles, Deep learning and machine learning, Intelligent control
Abstract: Effective thermal management strategy (TMS) plays a crucial role in ensuring the operational safety and energy efficiency of battery electric vehicles (BEVs), while also enhancing thermal comfort within the cabin. This study proposes a TMS based on proximal policy optimization (PPO) for integrated thermal management (ITM) system to improve temperature control and energy-saving potential. Considering the complex states and coupling relationships among subsystems within the ITM system, a hybrid attention mechanism is introduced, which simultaneously integrates physical features to enhance model generalization capability and convergence speed. Simulation results demonstrate that compared to rule-based TMS, the proposed improved PPO TMS reduces energy loss by 12.85% and effectively minimizes temperature fluctuations within the battery and cabin.

Keywords: Autonomous vehicles, Intelligent control
Abstract: Energy management for fuel cell electric vehicles (FCEVs) is a challenging trajectory optimization problem. Conventional studies primarily focus on trip-specific optimal control, where the power distribution is optimized based on a predicted finite-horizon driving profile. However, these methods often suffer from a limited look-ahead horizon and fail to guarantee long-run optimality within the stochastic traffic network where the vehicle operates. This study proposes a novel framework that integrates finite-horizon optimal control with traffic network-aware long-run average costs. We formulate the problem by embedding the long-run optimality, derived from network-level transition probabilities, into the terminal cost of the trip-specific optimization. This approach enables an adaptive target State of Charge (SOC) that aligns with global network efficiency while satisfying immediate driving constraints. Simulation results in a virtual traffic network demonstrate that the proposed integrated strategy consistently outperforms traditional trip-specific methods, achieving a maximum performance improvement of 11%. These findings highlight the necessity of network-level statistical awareness for maximizing the long-term energy efficiency of electrified mobility.

Keywords: Autonomous vehicles, Nonlinear control and applications, Artificial intelligence
Abstract: Real-time energy management strategies (EMS) are pivotal for optimizing the hydrogen economy and durability of fuel cell electric vehicles (FCEVs). While nonlinear model predictive control (NMPC) offers optimal performance, its heavy computational burden often hinders real-time implementation. Although data-driven methods can approximate NMPC behavior to accelerate computation, standard black-box neural networks often fail to guarantee physical consistency, potentially leading to infeasible control actions. To address these challenges, this paper proposes a physics-guided deep learning framework that approximates the NMPC control law. A hybrid CNN-LSTM architecture is constructed to extract local features from future driving conditions and capture temporal dependencies in battery state of charge (SOC) dynamics. Distinctively, a physics-informed loss function is formulated that explicitly incorporates the battery SOC differential equations and system operating boundaries into the training process. This approach enforces the network outputs to adhere to physical laws, thereby enhancing dynamic consistency and operational safety. Simulation results indicate that the proposed method achieves a computational speedup of two orders of magnitude compared to the numerical NMPC benchmark, satisfying real-time constraints. Furthermore, the physics-regularized policy maintains sub-optimal hydrogen economy.

Keywords: Control theories, Deep learning and machine learning, Intelligent control
Abstract: Decentralized stochastic optimization has emerged as a fundamental paradigm for large-scale machine learning. However, practical implementations often rely on biased gradient estimators arising from communication compression or inexact local oracles, which severely degrade convergence in the presence of data heterogeneity. To address the challenge, we propose Decentralized Momentum Tracking with Biased Gradients (Biased-DMT), a novel decentralized algorithm designed to operate reliably under biased gradient information. We establish a comprehensive convergence theory for Biased-DMT in nonconvex settings and show that it achieves linear speedup with respect to the number of agents. The theoretical analysis shows that Biased-DMT decouples the effects of network topology from data heterogeneity, enabling robust performance even in sparse communication networks. Notably, when the gradient oracle introduces only absolute bias, the proposed method eliminates the structural heterogeneity error and converges to the exact physical error floor. For the case of relative bias, we further characterize the convergence limit and show that the remaining error is an unavoidable physical consequence of locally injected noise. Extensive numerical experiments corroborate our theoretical analysis and demonstrate the practical effectiveness of Biased-DMT across a range of decentralized learning scenarios.

Keywords: Industrial applications, Artificial intelligence
Abstract: Robotic Mobile Fulfillment Systems (RMFS) involve tightly coupled decisions on order batching, station assignment, pod selection, and robot-to-delivery matching, making timely and cost-effective control difficult at scale. Although rolling-horizon optimization alleviates the computational burden of one-shot global planning, it can still induce myopic and weakly coordinated decisions across consecutive epochs. To address this issue, we develop an Structured Reinforcement Learning (SRL)-driven rolling-horizon framework for order selection and batching that exploits the global system state to improve release decisions over the fulfillment horizon. The actor evaluates candidate batches together with current resource occupancy and guides the up-level release of batch--station pairs. Given the released decisions, a lower-level combinatorial optimization model determines pod selection and robot-to-delivery matching. At each decision epoch, this low-level model is solved either by a Gurobi-based optimizer or by a fast level-by-level greedy constructor, providing a practical trade-off between solution quality and runtime. Experiments on multiple RMFS scales show that the proposed framework improves cumulative execution cost while maintaining online computational tractability.

Keywords: System identification and modelling
Abstract: This paper studies an adaptive control problem for simultaneous source localization and unmanned aerial vehicle (UAV) navigation based solely on noisy time-difference-of-arrival (TDOA) measurements obtained in cooperation with a fixed base station. An online localization algorithm is developed in which the source estimate is recursively updated using a stochastic gradient (SG) algorithm with a compensation term, requiring only the current measurement and the previous estimate at each step. Based on the localization estimate, an adaptive control law is constructed to steer the UAV toward an unknown stationary source. To reconcile the excitation requirements of localization with the stability requirements of navigation, a decaying dither signal is incorporated into the controller. It is shown that, under certain non-persistent excitation conditions, the localization algorithm converges asymptotically to the true source location and its convergence rate is explicitly characterized. Moreover, the closed-loop system is proved to satisfy the excitation condition required for localization, which guarantees that the UAV asymptotically reaches the source. Numerical simulations are provided to validate the theoretical results.

Keywords: System identification and modelling, Adaptive systems
Abstract: This paper investigates the estimation problems for stochastic large regression models under colored noise, where the system involves an infinite number of unknown parameters. To address this challenge, we design an online adaptive estimation algorithm that simultaneously estimates both the unobserved noise components and the unknown model parameters. The algorithm adaptively updates its dimensionality and computations as new data arrive incrementally, enabling dynamic learning adjustment without storing historical data. This design substantially reduces computational complexity and accelerates iterative updates. Furthermore, we rigorously establish the almost sure convergence of the estimation error without imposing independence, stationarity, or ergodicity assumptions on the regressors, making the results applicable even to systems with strongly correlated feedback. The effectiveness of the algorithm is validated through numerical simulations.

Keywords: Autonomous vehicles, Industrial applications
Abstract: Mixed traffic transportation systems are important as they allow different modes of human-driven vehicles (HDVs) and connected autonomous vehicles (CAVs) to share road space efficiently, improving accessibility and overall mobility while making better use of limited urban infrastructure. This paper investigates the problem of distributed estimation and tracking of HDVs by CAVs in mixed traffic setup. Each CAV is assumed to have access to local measurements/observations and to exchange state estimates and/or measurements with neighboring CAVs over a communication network. To enable cooperative inference, a distributed estimation protocol is designed in which each CAV updates its estimate using both local information and information received from its neighbors. Considering the notion of distributed observability for the CAV network, we define the condition under which the state of each HDV is observable to every CAV through data exchange over the network. Sufficient conditions on the network topology and estimator design are derived to guarantee distributed observability. In particular, it is shown that strong connectivity of the CAV communication graph, together with an appropriate choice of estimation gains, ensures that all HDV states can be asymptotically inferred by every CAV.

Keywords: Autonomous vehicles, Nonlinear control and applications
Abstract: This paper presents an experimental study on the robustness of linear model predictive control (MPC) for timeparameterized trajectory tracking in the Frenet frame. The focus is on operating-speed mismatch, where the prediction model is linearized at a nominal velocity while the vehicle operates at different speeds. A fixed linear MPC and a successive linearization MPC are implemented on a lab-scale autonomous vehicle platform using motion capture pose estimation and encoder-based velocity feedback. Experiments conducted over the speed range of 0.3–0.8 m/s demonstrate that fixed linear MPC maintains stable and accurate tracking despite moderate deviation from the nominal linearization point. While tracking errors increase at higher speeds, the degradation remains bounded and is primarily compensated by increased steering activity. A controlled comparison shows that successive linearization provides limited additional improvement in tracking accuracy while increasing control effort. The results indicate that fixed linear MPC offers a robust and computationally efficient solution for moderatespeed time-parameterized Frenet trajectory tracking in smallscale autonomous vehicle applications.

Keywords: Control theories, Intelligent control
Abstract: This paper addresses the optimal control problem for linear discrete-time systems with unknown system dynamics. By combining the Q-learning method in reinforcement learning with the transfer learning mechanism, this paper proposes a dynamic transfer Q-learning optimal control method driven by an advantage function. This method operates within an unsupervised reinforcement learning framework, using the recursive least squares method to update the parameters of the Q-function. Additionally, a dynamic transfer factor based on the advantage function is introduced to adaptively adjust the role of source task knowledge in the target task, thereby achieving optimal control. The paper proves the asymptotic stability of the proposed method under initial stable conditions. Simulation experiments verify the effectiveness of the proposed method.

Keywords: Autonomous vehicles, Cyber-physical systems and security, Control theories
Abstract: Autonomous ground vehicles (AGVs) rely on multiple heterogeneous sensing channels whose contributions to closed-loop stability are inherently unequal, making them vulnerable to selective integrity threats. This paper investigates channel-importance-driven selective measurement loss (SML) attacks, in which a resource-constrained adversary suppresses real-time updates of the most control-critical sensing channels and replaces them with stale values. By exploiting sensing heterogeneity, the proposed SML attack can significantly degrade tracking and stability performance while remaining covert. To counteract this threat, a defense control strategy is developed to adapt to the attack-induced measurement update pattern. Sufficient BMI-based conditions are established to guarantee closed-loop stability with prescribed H_infty performance under all admissible SML attacks. Simulation results on an AGV lateral dynamics model demonstrate that the proposed SML attack causes substantial performance degradation, whereas the proposed defense control effectively preserves stability and robustness under multi-channel attacks.

Keywords: Cyber-physical systems and security
Abstract: With the development of renewable energy technologies, the large-scale integration of renewable energy into urban power grids has posed significant challenges to power balance and operational security. Demand response (DR), as a flexible resource regulation strategy, can effectively improve the renewable energy absorption capacity of urban power grids and alleviate the peak-valley pressure of the power system. Therefore, this paper proposes a new collaborative optimization framework based on price-based demand response (PBDR) for urban power grids. Under this framework, a coordination and optimization dispatching model is constructed with the goal of minimizing the operating cost of the urban power grids. This model comprehensively considers factors including distribution network, distributed generation, price fluctuation risk, aiming to promote the consumption rate of renewable energy. Simultaneously, considering the stochasticity of renewable energy and load demand, this model is solved using the scenario tree method for multi-stage stochastic optimization. Finally, the method was implemented using historical data on an actual distribution network project in China. The results demonstrated that the established the method can effectively reduce the operating cost of urban power grids and significantly improve the utilization efficiency of renewable energy, verifying the accuracy and reliability of the proposed model.

Keywords: Artificial intelligence, Robotics and swarm intelligence, Deep learning and machine learning
Abstract: Multi-object tracking (MOT) in complex scenarios presents unique challenges due to frequent occlusions and similar appearances among objects. End-to-end methods have achieved remarkable progress by propagating object embeddings across frames for identity association. However, these methods suffer from observation dependency, where the learned embeddings primarily capture appearance and instantaneous position rather than motion dynamics. Consequently, association fails when visual input is degraded by occlusion. Inspired by how language models learn by predicting the next word, we find that such an autoregressive approach is ideally suited to address the lack of motion awareness in MOT. In this paper, we propose NextMOT, which learns motion-aware embeddings through autoregressive Next-State prediction. Specifically, we introduce a Temporal Autoregressive Predictor (TAP) to predict the next bounding box from historical object embeddings. This training-only module forces the embeddings to encode motion dynamics, enabling robust association under occlusion with zero inference overhead. Experiments on DanceTrack and SportsMOT demonstrate that NextMOT achieves state-of-the-art performance, with 70.4 and 73.9 HOTA, respectively.

Keywords: Intelligent control
Abstract: An extended state observer (ESO) is combined with an improved reinforcement learning (IRL) scheme to handle the heading control of a cargo ship. In standard reinforcement learning the long exploration time is a recurring problem, we mitigate this by placing the discretized states densely close to the setpoint and more sparsely away from it. The ESO supplies real time estimates of all ship states required by the learning agent, so the controller does not need an explicit ship model during operation. Numerical simulations indicate that the proposed scheme tracks the heading setpoint well, tolerates measurement noise, rejects external disturbances, and remains robust under variations of ship parameters.