State-of-the-art (SOTA) deep learning (DL) models may still produce incorrect inferences without signaling uncertainty, limiting their reliability in safety- and quality-critical automation systems. From a software testing perspective, runtime inference reliability assessment for modern convolutional neural networks and Transformer-based image recognition models remains challenging due to the lack of effective test oracles and the limited scalability of existing approaches. This paper proposes APPAREO, an unsupervised runtime reliability assessment framework for DL-based image recognition systems. Inspired by metamorphic testing (MT) principles, APPAREO constructs semantics-preserving transformed inputs and evaluates inference consistency between original and transformed samples. To strengthen the runtime test oracle, APPAREO further incorporates interpretation-consistency verification by comparing critical recognition regions extracted from heatmap-based visual interpretation techniques. Violations of prediction consistency or interpretation consistency are treated as indicators of unreliable inference behavior. We conduct extensive evaluations on large-scale image recognition tasks using convolutional-neural-network-based and Transformer-based models over the ImageNet and Caltech datasets. Experimental results show that APPAREO achieves an F1 score of 0.77 for unintended abnormal sample detection and outperforms several SOTA adversarial detectors with an average AUC of 0.80 under diverse adversarial attacks. In addition, we evaluate APPAREO in a real-world automated visual inspection scenario under both adversarial perturbations and environmental uncertainties. The results demonstrate that APPAREO is an effective and scalable runtime reliability assessment framework for modern DL-based image recognition systems.
Achieving carbon neutrality in mining operations requires energy-efficient control of open-pit mining trucks operating under complex conditions. Here, we present a multi-time-domain velocity and altitude prediction framework that integrates operation-mode recognition with bio-inspired optimization. We build a high-quality mine operation database through rigorous data preprocessing and develop a velocity prediction model using a Grey Wolf Optimizer-enhanced Radial Basis Function Neural Network (GWO-RBFNN), coupled with an altitude prediction model employing Plant Rhizome Growth Optimization (PRGO). For 5-second prediction horizons, the velocity model achieves a Root Mean Square Error (RMSE) of 1.7938 km/h—representing a 30.6% improvement over variants without operation-mode recognition and a 25.3% improvement over non-optimized alternatives—alongside a Mean Absolute Error (MAE) of 1.0996 km/h, a Mean Absolute Percentage Error (MAPE) of 9.6822%, a Nash-Sutcliffe Efficiency (NSE) of 0.9663, Willmott's Index (WI) of 0.9913, and Percent Bias (PBIAS) of 3.7154%. The altitude model achieves an RMSE of 1.5971 meters, with 47.7% and 43.0% improvements over ablation baselines, respectively; a MAE of 1.0832 meters; a MAPE of 0.3453%; an NSE of 0.9897; a WI of 0.9849; and a PBIAS of −0.0884%. Ablation experiments confirm the vital role of operation-mode recognition and optimization algorithms in improving predictive accuracy. Cross-mine tests show strong transferability across different geographic locations. This framework provides reliable feedforward data for energy management strategies in hybrid-electric mining trucks operating on engineered mine haul roads, where gradual grade transitions and operator anticipatory behavior allow current vehicle states to serve as proxies for near-term terrain prediction.
Fast transient response with a prescribed convergence time is essential for high-performance permanent magnet synchronous motor (PMSM) drives. Existing finite-gain prescribed-time (PT) results mainly use smooth Lyapunov analysis and cannot directly cover the non-smooth dynamics of super-twisting sliding mode (STSM) control. This paper develops a time-regularized practical PT stability framework using generalized derivatives for Filippov differential inclusions. A deterministic time-triggered regularization keeps the control gain bounded, while the framework guarantees semi-global convergence to a residual set within the prescribed time. Based on this result, an observer-free PT-STSM controller is designed for PMSM speed regulation. Unlike PT sliding mode control methods using discontinuous switching, the proposed controller generates a continuous control signal. Unlike conventional STSM, its convergence bound is independent of initial conditions. Experiments show that, compared with conventional STSM, the proposed method reduces overshoot by 34.9% and 26.0% at 900 and 1600 rpm, respectively, shortens the step-load recovery time by 39.7%, and reduces speed fluctuation under sinusoidal loading by 28.9%.
Accurate state estimation is crucial for power system security, yet challenges persist in data scarcity and multi-source heterogeneity. This paper presents a novel framework that integrates the Denoising Diffusion Implicit Model (DDIM), a deep generative architecture, with advanced data augmentation and multi-source fusion techniques. The proposed physics-informed method synthesizes measurement data via iterative denoising, effectively expanding the data manifold while preserving critical system constraints. Distinguishing itself from conventional generative approaches, the framework introduces a reverse-optimized diffusion process that removes Markovian dependencies and enables adaptive control over generation iterations. This architectural innovation leads to notable improvements in both computational efficiency and sample fidelity. To further enhance robustness, a new multi-source metrological fusion algorithm is developed to reconcile temporal and spectral discrepancies across heterogeneous data sources, including PMU and SCADA measurements. This ensures reliable and coherent data integration for state estimation. Building upon this, an enhanced Weighted Least Squares Measurement (WLSM) algorithm is formulated to estimate state variables with higher accuracy. Comprehensive empirical evaluations, including benchmark comparisons with contemporary generative models and IEEE system case studies, demonstrate the framework’s superior performance in sample generation quality, estimation accuracy, and computational efficiency. These results underscore the method’s effectiveness in addressing the challenges of multi-source data integration, marking a significant advancement in deep learning-driven power system state estimation.
This paper presents a mixed-integer linear programming formulation for a dynamic and heterogeneous multi-depot vehicle routing problem with time windows, which accounts for the following complexities: customers profitability and deadlines, skills requirements, shifting technician availability, flexible and partial scheduling, multi-day planning. Building on a gossip-based strategy, we propose a heuristic that: (i) assigns tasks to technicians based on skills compatibility and expected profitability; (ii) plans near-optimal routes within shift and time constraints; (iii) continuously updates assignments and routes in real-time (within minutes), to absorb stochastic delays, unforeseen impediments, and dynamically generated tasks; (iv) maintains private the real-time geolocation of the technicians; (v) is parallelizable in the sense that it can run simultaneously on multiple processors to speed up the execution. This paper is inspired by the real case study of DEDEM S.p.A., an international company that manages hundreds of technicians charged with refurbishing and repairing thousands of photo-booth machines spread across a wide geographic area, whose objective is to maximize the net profit by balancing per-task revenue against travel and labor costs. The performance of the proposed heuristic is demonstrated through numerical simulations over both synthetic data and real operational data provided by DEDEM S.p.A., saving up to 40% of profit loss and increasing up to 100% the net profit by efficiently handling stochastic delays.
This paper investigates the coded tracking control problem for nonlinear strict-feedback systems subject to denial-of-service (DoS) attacks and severe communication resource constraints. A high-gain observer is first constructed to ensure uniform ultimate boundedness of the state-estimation errors under DoS attacks. Next, a novel variable-bit event-triggered (VBET) encoding-decoding mechanism is proposed to encode the transmitted observer states. The codeword length is adjusted in real time based on both the recent DoS attack intensity and the normalized event-triggering error, thereby achieving a favorable trade-off between communication efficiency and attack resilience, while simultaneously determining the observer gain. To overcome the discontinuities in the decoded signal, a predictor-based continuous state reconstructor is then developed, which updates only at successful transmission instants and provides continuous state estimates. Subsequently, a backstepping controller is synthesized using these continuous estimates. Finally, a numerical example with several comparative simulations is provided to verify the effectiveness of the proposed approach.
3D Gaussian Splatting has emerged as a promising map representation that allows autonomous robots to localize themselves while reconstructing a photorealistic environment. However, its robustness in real-world deployments remains limited: the lack of effective noise mitigation and efficient loop closure leads to the accumulation of errors from motion blur and sensor noise. This paper presents μSLAM, a robust, real-time RGB-D Simultaneous Localization and Mapping (SLAM) system addressing these limitations through a synergistic, uncertainty-driven framework. It builds on a hybrid map integrating a sparse feature-based backbone with an uncertainty-aware Gaussian field. Local noise is mitigated through information-theoretic keyframe selection, coupled with a confidence-modulated adaptive optimizer for map refinement. Global consistency is ensured by a novel coarse-to-fine loop closure that leverages the embedded sparse map for robust initial alignment and the dense field’s rendering capabilities for high-precision refinement. Extensive experiments on public benchmarks and real-world robotic sequences demonstrate that μSLAM achieves state-of-the-art tracking accuracy and reconstruction quality, with superior robustness to noise and drift. Crucially, it is the only evaluated system delivering this level of performance while consistently operating in real-time (over 30 frames per second), making it a comprehensive solution for practical robotic applications.
Accurate estimation of knee joint torque from surface electromyography (sEMG) is important for exoskeleton control, but remains challenging because muscle dynamics are nonlinear and vary across participants. This study proposes a hybrid framework that fuses a Hill-type biomechanical model with a data-driven CNN–LSTM model for sEMG-based knee torque estimation. The Hill-type model provides physiological interpretability through participant-level calibration of selected activation and scaling parameters, whereas the CNN–LSTM model captures nonlinear spatiotemporal features from sEMG signals. A residual fusion strategy was used, in which the biome-chanical model served as the primary estimator and the data-driven model compensated for the residual error. Experiments with eight participants performing knee flexion–extension tasks showed that, relative to the model-derived reference torque, the fusion model reduced RMSE by 37.9% and 19.6% compared with the data-driven and biomechanical models, respectively. The proposed approach combines interpretability and adaptability, providing a feasible solution for personalized torque estimation in wearable robotic systems.
Human walking and joint-angle trajectories are complex and highly individual-specific. The study presents a compact wavelet-domain approach that predicts joint movement from body measurements and reconstructs full trajectories using an inverse wavelet transform. Lower-limb joint signals, including those from the hip, knee, and ankle, are sampled at 1,000 frames per stride and processed into a 15-coefficient format using a discrete wavelet transform (DWT). A feedforward neural network connects body measurements to these coefficients. The predicted coefficients are converted back using inverse DWT (IDWT) to generate 1,000 joint-angle trajectories per frame. The proposed framework was evaluated on a publicly available motion capture dataset comprising 138 able-bodied adults across the lifespan. Then, the mean squared error (MSE) and Pearson correlation coefficient were evaluated between the original and reconstructed signals. The proposed WaveGaitNet framework achieves R2 values of 0.97, 0.96, and 0.88 for hip, knee, and ankle joints, respectively, with corresponding RMSE values of 3.01°, 4.85°, and 2.67°. This demonstrates improved reliability in reconstructing individual joint-angle trajectories using only basic body measurements.
Efficient and robust trajectory tracking control is crucial for autonomous agricultural vehicles (AAVs) operating under complex and noisy field environments. However, measurement noises, model uncertainties, and external disturbances severely degrade tracking accuracy in practice. To address these challenges, we propose an augmented observer-based nonsmooth integral trajectory tracking control scheme for AAVs based on a mixed-offset model. The framework integrates an augmented fixed-time extended state observer (AFESO) and a generalized proportional integral observer (GPIO) to handle the longitudinal and lateral–heading subsystems, respectively. The AFESO achieves fixed-time disturbance estimation with built-in noise attenuation, while the GPIO ensures accurate disturbance compensation. On this basis, a nonsmooth integral controller is designed to guarantee fast convergence and smooth actuator behavior, effectively mitigating chattering compared with conventional sliding mode-like methods. Rigorous stability analysis is conducted via the fixed-time Lyapunov theory to prove closed-loop convergence. Finally, both simulation and field experiments demonstrate that the proposed control scheme significantly improves tracking accuracy and robustness over existing approaches.
This paper proposes a physics-based adaptive Koopman Model Predictive Control (MPC) strategy for combined spacecraft attitude stabilization under inertia uncertainties and active target maneuverability. A novel, quaternion-based Koopman model is constructed from a set of analytical lifting functions derived from the quaternion kinematics, which provides a more compact and physically interpretable linear representation of the nonlinear dynamics compared with the conventional black-box EDMD and higher-dimensional DCM-based model. Leveraging the linear structure of this nominal model, a gradient descent-based update law is employed to efficiently identify time-varying inertial uncertainties from real-time input/output data. By integrating this adaptive linear model into the MPC framework, the optimal control problem reduces to a computationally efficient Quadratic Program (QP), thereby significantly lowering the online computational burden compared to nonlinear adaptive MPC. Recursive feasibility and regional input-to-state stability are formally established through the design of terminal ingredients for the MPC. The effectiveness and superiority of the proposed strategy are validated through comparative simulations of an attitude stabilization task for combined spacecraft in a high-fidelity 3D simulator.
Offline imitation learning (IL) is a practical route to embodied policy learning when online exploration is unsafe, expensive, or unavailable. Its main limitation is the closed-loop covariate shift: a policy trained only on fixed demonstrations can enter states outside the expert support, where behavioral cloning provides no recovery supervision and small errors compound. We propose Synthetic Imitation Learning (SynIL), an offline IL framework that augments demonstrations with synthetic recovery trajectories rather than only filtering or reweighting logged data. SynIL uses generative dynamics models to enlarge the training distribution around failure-prone states and retains transitions whose successor states move back toward the expert state manifold. In low-dimensional benchmarks, SynIL trains a lightweight latent world model to imagine short-horizon rollouts from suboptimal states and uses a positive-unlabeled state discriminator to assign proxy rewards without access to task rewards. In high-dimensional robot manipulation, SynIL instantiates the same principle with a training-free visual synthesis pipeline that combines pretrained video generation, segmentation, point tracking, and monocular depth estimation to recover imageaction trajectories. Experiments on 15 D4RL tasks show that SynIL achieves first- or second-place performance on 8 tasks across navigation, locomotion, and manipulation. Real-world UR5 experiments improve the average success rate by 17.5% over behavior cloning, indicating that synthetic recovery data can improve robustness without online environment interaction.
This paper addresses the load frequency control (LFC) problem for electric vehicle (EV)-integrated power systems under state and input constraints. A decentralized adaptive dynamic programming (ADP) control scheme with an event-triggered (ET) mechanism is developed to achieve efficient frequency regulation across interconnected areas. To enforce state constraints, a logarithmic barrier transformation is introduced. Furthermore, a non-quadratic utility function is used to handle input constraints. Additionally, an event-triggered rule with a dead-zone operator is introduced to reduce communication and controller updates while guaranteeing the absence of Zeno behavior. The uniform ultimate boundedness of the closed-loop system is established using Lyapunov theory. Simulation results demonstrate the effectiveness of the proposed algorithm
Efficient production and rapid, sustainable prototyping of large metal components are key in industries such as aerospace, energy, oil and gas, maritime, and automotive. Wire Arc Additive Manufacturing (WAAM) is characterised by high deposition rates, material efficiency, and suitability for producing large-scale metal components with strong mechanical properties. In WAAM, reliable in-process monitoring and control of material deposition are essential for geometric accuracy and final-part quality. In this paper, we address the problem of inline measurement of the weld bead width using a cost-constrained, low-specification vision system, where plasma saturation and camera limitations prevent direct observation of the weld pool. We introduce a visual measurement and validation framework that, unlike conventional weld-pool-based approaches, shifts the measurement target from the plasma-occluded weld pool to the uppermost deposited layer. By combining temporal layer-mask propagation with adaptive local geometric analysis, our method separates the deposited layer from the underlying weld seam and provides stable estimates of the weld bead width. The performance of our method has been tested under realistic application scenarios, closely emulating actual manufacturing conditions. Additionally, we introduce the LAVA-WAAM, a new segmentation dataset, comprising 7.8 hours of densely annotated video collected across multiple experimental sessions. Measurement accuracy is quantitatively validated by comparing the in-process bead-width estimates with reference measurements obtained from 3D scans of the completed parts. The reported validation results show an RMSE of approximately 0.12 mm relative to post-build 3D scans, demonstrating the applicability of the proposed framework under challenging WAAM conditions.
Contouring accuracy stands as the paramount performance metric for H-type gantry platforms, indispensable for advanced equipment manufacturing. However, under high-dynamic conditions, various uncertain factors such as inertial variations, nonlinear friction, external disturbances, flexible structural vibrations, and parameter uncertainty are significantly amplified. These severely degrade contouring performance and even threaten system stability. To address these critical challenges, this paper proposes an adaptive fractional-order integral terminal synergetic control scheme rooted in synergetic theory and immersion and invariance (I&I) theory. Compared with other advanced robust control strategies, the proposed approach effectively overcomes the severe chattering problem arising under such demanding conditions. By innovatively incorporating fractional-order integral terms into the macro variable design, the system is endowed with augmented robustness and additional degrees of freedom for regulating error dynamics. This significantly enhances the capability of the system to resist a wide range of uncertainties under high-dynamic conditions. Furthermore, an I&I-based parameter identification mechanism is proposed for the H-type gantry platform system, which ensures asymptotic convergence of parameter estimation errors rather than only guaranteeing bounded error stability. Finally, following a rigorous theoretical stability analysis, the experiments on the H-type gantry platform are conducted to verify the effectiveness and demonstrate the superiority of the proposed method.
Strip steel surface defect detection (S3D) is a critical task in industrial quality control, and existing detectors often struggle to precisely localize fine defect structures within complex textured backgrounds. High-frequency cues, such as subtle edges and texture patterns, are known to be effective for capturing fine details; however, they are often progressively attenuated or even lost in deep convolutional backbones due to their intrinsic low-pass filtering behavior. To address this limitation, we propose a high-frequency guided feature aggregation (HFFA) framework that explicitly decouples low-frequency semantic information and high-frequency defect-sensitive cues into two complementary representation pathways. Specifically, the backbone network is explicitly interpreted as a low-frequency semantic extractor for modeling global context and defect semantics, while a Laplacian feature pyramid network (LFPN) is introduced in parallel to explicitly extract multiscale high-frequency structural features by removing low-frequency components. Based on this frequency-decoupled representation, we design high-frequency injection modules (HFIMs) that use fixed high-frequency priors to modulate backbone features, selectively enhancing defectrelevant regions while suppressing noise. Finally, we develop a routing enhancement module (REM) that enable content-adaptive multiscale collaboration through dynamic spatial routing, improving the adaptability of fused representations to varying defect scales. Experimental evaluation on three benchmark datasets for S3D shows that HFFA consistently outperforms twenty-three SOTA methods. Furthermore, cross-domain evaluation on pipeline weld defect detection confirms its generalization, while integration of its key high-frequency extraction and injection components into the YOLO series substantially improves their prediction accuracy, confirming the robustness and transferability of our approach. Our code is available at https://github.com/hpguo1982/HFFA.
Open-world instance classification in LiDAR point clouds aims to discover and categorize instances of unknown classes unseen during training, without access to their ground-truth labels. The task is extremely challenging since only knownclass labels are available for training. Existing LiDAR classification methods often depend on restrictive priors, including a predefined number of novel classes or additional annotations for unseen objects, which limits their applicability in continually changing environments. In this paper, we propose LOIC, an open-world instance classification framework for LiDAR point clouds with instance-level contrastive learning. LOIC first obtains candidate unknown instances and their voxel features through open-set semantic segmentation and ellipsoidal clustering. The introduced Classification Feature Extraction Head module (CFEH) then transforms point-wise features into discriminative cluster-based instance representations. The proposed Classification with Geometry-aware High-dimensional Features module (CGHF) further integrates orientation-robust geometric cues with the learned representations. Unknown instances are continuously assigned to discovered categories or initialized as novel categories according to their distances from dynamically maintained class prototypes, without requiring the unknown-class number in advance. We establish an evaluation benchmark on the nuScenes, SemanticKITTI, and in-house Campus datasets. Experiments show that LOIC outperforms state-of-the-art approaches adapted from the visual and feature-clustering domains, achieving improvements by 3.0, 4.8, and 7.9 percentage points in PQ, mIoUu, and mPreu, respectively. The source code of our method will be publicly available at https://github.com/nubot-nudt/LOIC.
This paper proposes a distributed adaptive optimal fault-tolerant formation containment control scheme for high-order nonlinear multi-agent systems (MASs) subject to actuator faults and unknown dynamics. Radial basis neural networks are employed to approximate unknown nonlinear functions. Subsequently, a distributed extended observer is designed to estimate unmeasurable system states. Based on the fully actuated system (FAS) approach, an adaptive reinforcement learning algorithm is developed within an identifier–critic–actor framework to handle high-order dynamics and derive the control input, thereby achieving nearly optimal control. Moreover, an (n-1)-order sliding mode surface is constructed to regulate the tracking error, and an adaptive optimal fault-tolerant control strategy is developed. Theoretical analysis shows that even in the presence of actuator faults, MASs can maintain their output within the convex hull formed by the leaders, while achieving high precision formation trajectory tracking and optimized overall performance. Finally, the feasibility and effectiveness of the proposed control scheme are verified through a simulation example.
Developable surfaces are widely used in industrial applications such as sheet metal forming and surface processing, where robots are often introduced to assist with various tasks. In these applications, especially when relying on demonstration in human-machine collaboration to perform complex tasks, ensuring that robots can accurately follow curved surfaces presents a significant challenge. Traditional methods typically employ pointwise mapping to project demonstration paths from a plane onto a curved surface, due to the ease of demonstrating actions on a flat surface. However, robotic motions are typically encoded as continuous state policy such as dynamic movement primitives rather than discrete path points, which complicates the transfer of demonstrated motion to geometrically constrained surfaces. In response to this challenge, we model the robot’s working surface as a manifold and exploit the fact that developable surfaces are intrinsically isometric to planar. Building on this, we propose a method that enables smooth and consistent skill transfer from planar demonstrations to curved surfaces while preserving geometric characteristics of motion. Experimental validation, including physical experiments on UR5 robot laser marking tasks and numerical simulations on several representative curved surfaces, demonstrates the effectiveness of our approach, highlighting its potential for broader robotic applications involving constrained surfaces.
