PurposeNon-standard structural components present substantial challenges for robotic welding path planning. This study aims to propose an efficient method for weld feature extraction that combines lightweight instance segmentation with point cloud analysis to overcome these challenges.Design/methodology/approachThe method uses a You Only Look Once-Fast (YOLO-Fast) model to segment workpieces and map 2D image pixels to 3D point cloud data, enabling precise extraction of regions of interest. A random sample consensus-oriented bounding box algorithm is then employed to localize weld feature points accurately.FindingsYOLO-Fast reduces model parameters by 39.08%, lowers computational cost by 29.15% and increases inference speed by 17.17%, while maintaining high segmentation accuracy. Experiments on tower base structures achieve a root mean square error (RMSE) below 2.98 mm, improving extraction accuracy by at least 15.4%. The RMSE for inner and outer weld features reaches 1.96 mm, demonstrating the method's robustness and practical applicability.Originality/valueThis study introduces a novel approach for efficient and precise weld feature extraction in complex, nonstandard structural components, providing a foundation for enhanced robotic welding path planning in challenging industrial scenarios.

High-penetration renewable integration and the spatial mismatch between energy supply and demand increase the complexity of multi-regional power system planning. Existing AI techniques in planning include renewable forecasting and demand prediction. These methods enhance the input fidelity of Mixed-Integer Linear Programming (MILP) models. However, they mainly focus on explicit and observable parameters. Multi regional planning also depends on latent and unobservable inter-regional responses. Such latent features are difficult to capture in the MILP pipeline. To address these limitations, we propose the Constraint-Adaptive Graph-Convolution (CAG-C) Reinforcement Learning (RL) model, which performs end-to end policy optimization. The planning problem is formulated as a finite-horizon Markov Decision Process, where the agent iteratively adjusts regional power exchange strategies. A graph convolutional network encodes spatial and resource correlations, while a constraint-adaptive architecture transforms raw actions into feasible plans via region-wise scaling, power balance correction, and cost-optimal transmission allocation. This modular design decomposes the high-dimensional planning task into tractable subproblems, ensuring operational feasibility while enabling efficient learning. Numerical results demonstrate that the proposed framework is both highly extensible and interpretable. As training progresses, the evolving solution trajectory reveals valuable insights into each region's functional role, offering a promising new paradigm for future power system planning.

Pneumatic separation can exhibit unstable performance when the feed composition fluctuates while operating parameters remain fixed. This work investigates a perception-informed airflow regulation approach, demonstrated on a representative fibrous–granular mixture case study. We propose LGDNet, a lightweight visual ratio estimation network (0.08 M parameters) built with Ghost-based operations and learned grouped channel convolution (LGCC), to estimate mixture composition from dense images. A dedicated 21-class dataset (0–100% in 5% increments) containing approximately 21,000 augmented images was constructed for training and evaluation. LGDNet achieves a Top-1 accuracy of 66.86%, an interval accuracy of 74.10% within a ±5% tolerance, and an MAE of 4.85, with an average inference latency of 28.25 ms per image under the unified benchmark settings. To assess the regulation mechanism, a coupled CFD–DEM simulation model of a zigzag air classifier was built and used to compare a regime-dependent airflow policy with a fixed-velocity baseline under representative prescribed inlet ratios. Under high impurity loading (r=70%), the dynamic policy improves product purity by approximately 1.5 percentage points in simulation. Together, the real-image perception evaluation and the mechanism-level simulation study suggest the feasibility of using visual ratio estimation to inform airflow adjustment; broader generalization and further on-site validation on real equipment will be pursued in future work.

Hemiparetic gait is characterized by inter-limb asymmetry, which makes reference trajectory generation difficult for unilateral lower-limb exoskeletons. This study presents a healthy-limb-driven contralateral gait prediction framework integrated into a single-leg exoskeleton platform under a strictly unilateral sensing configuration. The proposed system combines wearable inertial sensing with a Seq2Seq model featuring a Bi-LSTM encoder and Luong attention to establish a closed-loop reference generation and execution pipeline. The prediction model is trained and evaluated using joint-angle data collected from healthy subjects under controlled walking conditions. Experimental results demonstrate stable contralateral hip and knee trajectory reconstruction with improved accuracy compared with baseline LSTM-based methods. System-level validation through overground walking experiments confirms that the generated reference trajectories can be executed online within the mechanical and control constraints of the developed platform. These findings support the feasibility of unilateral data-driven reference generation for lower-limb exoskeleton control.

To enhance the accuracy and stability of industrial robot teaching, we propose an improved teaching gun that incorporates an actively illuminated rectangular light pattern. The system employs binocular vision to extract the edge features of the light pattern and to establish a dedicated light-pattern coordinate system. Two transformation matrices are computed: one that maps the light pattern coordinate system to the binocular camera coordinate system, and another that maps the teaching point coordinate system to the light pattern coordinate system. By combining these transformations, the spatial pose of the teaching point relative to the camera is determined, enabling precise robot teaching. Experimental results demonstrate that the proposed method achieves pose accuracy sufficient to meet the requirements of robot teaching, with average position errors of 0.452 mm, 0.325 mm, and 0.424 mm in the x, y, and z directions, and average orientation errors of 0.810 deg, 0.308 deg, and 0.537 deg. Furthermore, the system exhibits excellent repeatability, with an average error of 0.073 mm, meeting the accuracy and stability requirements of industrial teaching applications.

Unilateral gait impairment after hemiparesis is characterized by pronounced asymmetry between limbs, which challenges trajectory generation in unilateral lower limb exoskeletons. Conventional mirroring strategies are limited in this setting and may transfer pathological compensations. This study uses healthy limb motion and a Temporal Convolutional Network with Attention (TCN-Attention) to reconstruct contralateral joint kinematics under causal and unilateral sensing constraints. In healthy subjects, the model achieved hip RMSE of $6.342^{\circ }$, knee RMSE of $6.813^{\circ }$, correlations above 0.90, and lower RMSE/MAE than recurrent baselines in the main comparison. A preliminary offline feasibility evaluation on six individuals after stroke showed temporal synchronization without individual fine tuning (phase error $\approx 10\%\text{--}12\%$). Because amplitude deviations remained in pathological gait, the predicted trajectories were treated as candidate references derived from healthy training patterns for supervised rehabilitation control.

The advancement of intelligent monitoring for Gas Metal Arc Welding (GMAW) is critically constrained by the scarcity of high-fidelity labeled data and the prohibitive cost of manual annotation. To address this challenge, a Physics-Guided Data Programming (PGDP) framework is proposed to integrate domain-specific welding physics with weak supervision for scalable arc stability monitoring in robotic weaving GMAW. Instead of relying on manual labels, the framework leverages expert knowledge, calibrated via synchronized high-speed imaging, to encode multi-dimensional physical heuristics into programmatic Labeling Functions (LFs). A generative model then aggregates these potentially conflicting weak supervision signals to produce probabilistic training labels for a discriminative deep learning model. SHapley Additive exPlanations (SHAP)-based interpretability is used for feature optimization and model analysis. It reveals that high-order physical features, such as Power Kurtosis, serve as informative signatures of arc instability and are difficult to capture using conventional threshold-based criteria. Validated on a robotic weaving V-groove welding dataset, the proposed method consistently outperforms representative statistical and learning-based baselines, achieving an F1-score of 0.952 and demonstrating robust generalization across diverse operating conditions. Overall, this study establishes a cost-effective and data-efficient paradigm for industrial welding quality monitoring by systematically integrating physics-guided weak supervision with scalable data-driven learning.

Human gait exhibits stable contralateral coupling, making healthy-side motion a viable predictor for affected-limb kinematics. Leveraging this property, this study develops FusionTCN–Attention, a causality-preserving temporal model designed to forecast contralateral hip and knee trajectories from unilateral IMU measurements. The model integrates dilated temporal convolutions with a lightweight attention mechanism to enhance feature representation while maintaining strict real-time causality. Evaluated on twenty-one subjects, the method achieves hip and knee RMSEs of 5.71° and 7.43°, correlation coefficients over 0.9, and a deterministic phase lag of 14.56 ms, consistently outperforming conventional sequence models including Seq2Seq and causal Transformers. These results demonstrate that unilateral IMU sensing supports low-latency, stable prediction, thereby establishing a control-oriented methodological basis for unilateral prediction as a necessary engineering prerequisite for future hemiparetic exoskeleton applications.

To improve the absolute positioning accuracy of a six-degree-of-freedom (6-DOF) industrial robot, this study presents a method for predicting and compensating non-geometric errors using an optimally pruned extreme learning machine (OP-ELM), optimized by an enhanced backtracking search algorithm, alongside geometric parameter errors calibration. The proposed approach maps position errors of target points in Cartesian space to joint space and enhances the mutation strategy of the backtracking search algorithm to optimize the construction of an OP-ELM neural network model for joint space error prediction. This methodology enables the prediction and compensation of non-geometric errors for target point positions within the robot’s workspace. Experimental validation of the method was conducted using a 6-DOF industrial robot, demonstrating that the average absolute position error of the robot’s end-effector center point decreased from 1.229 mm to 0.164 mm, with the maximum absolute position error reduced from 1.938 mm to 0.275 mm. These results confirm the accuracy and efficacy of the proposed approach.

Rotational offsets, oil reflections, and noise interference pose significant challenges to cylinder heads identifying and localizing. In this article, we propose a recognition and localization method that integrates oriented FAST (oFAST), speeded-up robust features (SURFs), and grid-based motion statistics (GMSs) with minimum surrounding contours. First, the oFAST extracts feature points from the cylinder head image, and a 64-D SURF descriptor is generated for these feature points to obtain initial matching results. Second, the GMS algorithm performs grid segmentation and eliminates the mismatched points. Then, an adaptive threshold key point evaluation method is applied to refine the feature points and eliminate residual edges. Finally, the pose information of the cylinder head is obtained by background suppression and minimum surrounding contour (BSMSC) method. We conduct cylinder head recognition and localization experiments on the self-developed sorting robot experimental platform. Experimental results show that the OFS-GMS matching recognition algorithm has an average recognition accuracy of 99%. The BSMSC-based pose estimation method operates with an average processing time of 48.14 ms and a maximum angular error of less than 1.5 degrees.

In ultra-precision polishing tests, due to the corrosive and adhesive properties of the polishing abrasive, the spray system faces wear, blockage and oxidation problems. To solve these problems, this paper studied nozzles and verified the wear mechanism of the coated and uncoated nozzles by simulating the operating conditions after assembling the spray system. In the early stages of the experiment, the polishing speed of the spray system (v = 10 L/min), the feed rate (vf = 7.8 mm/min) and the polishing pressure (2~3.5 MPa) were maintained. The wear mechanism and surface morphology features of the nozzles in each case were analyzed by Hitachi S-3400N electron microscopy. When comparing the surface morphology of the nozzle coated with titanium alloy and the uncoated one, the results show that there is a significant difference in the corrosion resistance of the coatings to the abrasive particles. A significant effect was seen on the wear morphology, proving that the nozzle wear mechanism includes wear, adhesion and diffusion. Under the experimental conditions of a lateral velocity of 7.8 mm/min and a polishing force of 2 MPa, BK7 was polished using nozzles 1 and 2, resulting in a surface roughness of 75 nm and 35 nm while PV values were 125 nm and 67 nm, respectively. The excellent quality of nozzle 2 (coating nozzle) was proven, further demonstrating the superiority of the coating nozzle. Finally, the lifespan of the nozzle was extended and the surface accuracy of BK7 was improved by coating titanium alloy composite material on the 304 stainless steel nozzle.

The widespread use of lithium-ion batteries has somewhat alleviated the fossil energy crisis. However, when the state of health (SOH) of a lithium-ion battery drops to 80 %, it usually needs to be taken out of service from its original workplace. Accurately evaluating the state of energy (SOE) of retired batteries can effectively avoid unnecessary losses due to accidental failures of retired lithium-ion batteries in other application scenarios. Therefore, this paper proposes to come to the SOE estimation model of retired lithium batteries based on reinforcement learning Q-learning. Firstly, according to the unstable characteristics of retired Li-ion batteries, Qlearning is used to optimize the weight parameters of three algorithms, namely, back-propagation neural network (BPNN), long and short-term memory network (LSTM) and support vector regression (SVR), to establish the Q-BPNN-LSTM-SVR (QBLS) integration algorithm. Then, considering that the three factors of temperature, voltage, and current affect the state of energy of retired lithium-ion batteries, they are selected as inputs to the QBLS algorithm. Finally, the QBLS algorithm estimates the SOE of individual retired batteries and retired battery packs at different temperatures and operating conditions. The results demonstrate that the method has high precision in estimating the SOE of retired Li-ion batteries, with a Max AE less of than 3 %, and the maximum values of MAE and RMSE do not exceed 0.70% and 0.80%, respectively. The method above offers a dependable foundation for using and administering retired lithium-ion batteries.

In actual welding environments, factors such as the welding process, light intensity, material properties of the workpiece, and surface quality introduce interferences such as noise, reflection, curve around, fume, and splash in laser images, which degrade the quality of centerline extraction and the precision of weld seam feature point identification. To address these challenges, a comprehensive method for precisely extracting weld seam feature point in multi-interference environments is proposed. Initially, an adaptive thresholding method based on the minimum error prior values (AT-MEPV) is proposed for image binarization to separate laser stripe information from background noise. Subsequently, a novel method for laser stripe centerline extraction is proposed, combining the grayscale centroid method with the Levenberg-Marquardt dual-threshold center point prediction method, to accurately extract the centerline of the laser stripe. Finally, an adaptive iterative random sample consensus algorithm is proposed to extract weld seam feature points precisely. Experimental results demonstrate that the proposed method can effectively extract feature points from various types of weld seams under single and multiple interference conditions, including T-weld, short side T-weld, and V-weld. The average extraction errors for these weld types are 2.1177 pixels, 2.2021 pixels, and 1.4810 pixels, respectively, with corresponding root mean square errors of 0.0544 pixels, 0.0535 pixels, and 0.0711 pixels.

In the field of abrasive-water-jet polishing technology, the influence of nozzle geometry on nozzle wear and internal-structure erosion in abrasive-water-jet polishing technology is studied, and the nozzle design is optimized through experiments and a numerical simulation to improve the stability and efficiency of the abrasive jet. The mathematical model between the cross-sectional area of the nozzle and the dimensionless length of the nozzle is established, as well as the variation in the static pressure of the nozzle and the length of the nozzle. Through Fluent simulation, it is found that when the nozzle length is 12 mm, the abrasive-phase acceleration is sufficient and the erosion intensity is minimal. After 480 h of erosion experiments, the erosion profile of nozzle cavity was detected. The results show that the erosion strength of the 12 mm nozzle is the least, followed by the 6 mm nozzle, and the 18 mm nozzle is the strongest, which is consistent with the simulation results.

The accurate establishment of battery model can improve the design reliability and reduce the design risk, which provides an important basis for the research of battery. First, the key parameters of the Li-ion battery model are identified by the least square method based on the full-battery equivalent circuit model of the single-particle impedance spectrum, and the diffusion coefficient and exchange current density under different temperatures and SOC conditions are calculated. At the same time, the one-dimensional thermal rate model is used as the heat source of the three-dimensional model, and the mean temperature T of the three-dimensional model is calculated by using Fourier's law, and T is fed back to the one-dimensional model as the key parameter to modify the conductivity, diffusion coefficient, and exchange current density, and a semi-empirical electrochemical-thermal coupling model with two-factor parameter modification is established. Finally, the model is verified by the temperature field distribution and discharge voltage curve at different discharge rates. The maximum temperature difference is less than 3.1 degrees C, and the maximum voltage difference error is less than 0.131 V. The results show that the improved model can accurately reflect the influence of temperature on the model parameters, and has high accuracy in the estimation of battery terminal voltage and SOC.

This paper proposes a novel active disturbance rejection control (ADRC) scheme to achieve the spacecraft relative pose synchronization in the libration orbits under the constrations of the actuator saturation and time-delays. The relative kinematics and dynamics between two spacecraft are modeled using dual quaternions to avoid singularity. A set of dimensionless variables is introduced to improve computational accuracy. An extended state observer, consisting of a time-delay predictor and an anti-windup compensator, is designed to address actuator saturation and communication time delays. A composite controller is then proposed, based on the designed observer, to mitigate the impact of total disturbances. Finally, two illustrative numerical simulations, including a spacecraft close rendezvous and docking scenario and a formation flying scenario, are provided to verify the effectiveness of the proposed ADRC scheme. (c) 2024 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

Robotic automated welding of non-standard steel structures presents significant challenges, particularly for electric power tower bases. This study introduces a novel approach that integrates the You Only Look Once-Compact Invert Block and Efficient Local Attention (YOLO-CE) model, an enhanced version of YOLOV8 for 2D image segmentation, with 3D point cloud technology. The YOLO-CE model is used to accurately extract point cloud data from the target area, which is then processed using the MSAC algorithm for efficient plane segmentation. Weld lines are identified through plane equations, allowing for initial weld point cloud extraction. To further refine accuracy, an optimized evaluation equation is developed that accounts for both the distance between the weld point cloud and the fitted plane, and the angle between their normal vectors. This enables precise classification of the weld point cloud. From this classification, key weld feature points are identified, and their exact positions are determined by calculating the distances between these points and their intersections with three planes. The reliability of the proposed method was validated using a robot for precise measurements, with a total error margin of less than 1.5084 mm, demonstrating high accuracy and stability. Post-operation inspections confirmed that the welds were filled and free from defects, meeting all process requirements. The YOLO-CE model achieved a mIoU of 96.38% and a precision of 99.8%, highlighting its effectiveness. This method provides an efficient and precise solution for the automated welding of non-standard steel structural components and has promising application potential.

Accurate and effective battery state-of-health (SoH) monitoring is significant to guarantee the security and dependability of electrical equipment. However, adapting SoH estimation methods to diverse battery kinds and operating conditions is a challenge because of the intricate deterioration mechanisms of batteries. To solve the issue, in this article, a novel multi-input metabolic long short-term memory (MM-LSTM) framework is developed. A degradation state model is created with the LSTM network to describe the intricate deterioration mechanisms. To convey more information about battery aging, the capacity degradation, sample entropy of discharge voltage, and ohmic internal resistance increment are extracted as the inputs of the model. To estimate SoH with a few data, the metabolic mechanisms are introduced to update the inputs and reflect the latest developments in aging. The accuracy and robustness of the proposed MM-LSTM framework are verified in different aspects using two kinds of batteries, and the maximum estimation error of SoH is within 1.98%. The findings indicate that the MM-LSTM framework implements the transfer application of SoH estimate successfully, and the framework’s versatility has been proven.

With the increasing adoption of lithium-ion batteries in energy storage systems, accurately monitoring the State of Health (SoH) of retired batteries has become a pivotal technology for ensuring their safe utilization and maximizing their economic value. In response to this need, this paper presents a highly efficient estimation model based on the multi-input metabolic gated recurrent unit (MM-GRU). The model leverages constant-current charging time, charging current area, and the 1800 s voltage drop as input features and dynamically updates these features through a metabolic mechanism. It requires only four cycles of historical data to reliably predict the SoH of subsequent cycles. Experimental validation conducted on retired Samsung and Panasonic battery cells and packs under constant-current and dynamic operating conditions demonstrates that the MM-GRU model effectively tracks SoH degradation trajectories, achieving a root mean square error of less than 1.2% and a mean absolute error of less than 1%. Compared to traditional machine learning algorithms such as SVM, BPNN, and GRU, the MM-GRU model delivers superior estimation accuracy and generalization performance. The findings suggest that the MM-GRU model not only significantly enhances the breadth and precision of SoH monitoring for retired batteries but also offers robust technical support for their safe deployment and asset optimization in energy storage systems.