Abstract:Convolutional neural network (CNN) have been limited in their ability to effectively model long-range dependencies due to their localized convolution operations. In contrast, vision Transformer achieves explicit modeling of global dependencies through mechanisms such as self-attention. In surface defect detection tasks, especially in scenarios with complex background textures or diverse defect morphologies, vision Transformer shows superior performance compared with CNN. This article provides a comprehensive review of recent domestic and international research progress and challenges in surface defect detection based on vision Transformer, focusing on two dimensions: The technical advantages and application methodologies, as well as key challenges and corresponding strategies. Firstly, the fundamental definition of surface defect detection is elucidated, and the technical characteristics and main challenges in this field are summarized. Secondly, the technical advantages and key challenges of the vision Transformer in the context of defect detection are analyzed. Subsequently, leveraging the technical strengths of vision Transformer, typical applications in surface defect detection tasks are examined in detail, including handling complex texture background interference, achieving multimodal information fusion, and integrating local-global feature information based on a modular design approach. Subsequently, the article discusses the main optimization strategies and solutions adopted by vision Transformer to address key challenges in surface defect detection, such as scarce sample data, high computational complexity, insufficient real-time performance, low training efficiency, and poor performance in detecting small targets. Finally, future research directions and development trends of vision Transformer in the field of surface defect detection are prospected, such as the development of transfer learning-based pre-trained models and their advanced fusion with multimodal methodologies, among others.

Abstract:Aiming at the problem of limited recognition accuracy caused by the loss of shallow features in the detection of micro cracks on plate heat exchanger surfaces, this paper proposes a detection method based on a lightweight dense convolutional network (DenseNet) architecture. As a single-category, multi-scale distributed cluster of linear small objects, the core challenge of microcrack detection lies in the effective learning and preservation of spatial detail features. The main contributions are as follows: First, a theoretical model matching the receptive field to defect size is established, with mathematical formulas derived for hierarchical configuration, followed by model instantiation based on the characteristics of corrugated plate defects in real industrial scenarios. Second, a detail enhancement mechanism is designed, which preserves critical spatial features by disabling downsampling operations and progressively expands the receptive field through stacked 3×3 small convolutional kernels, effectively balancing feature resolution and semantic abstraction. Finally, a defect instancelevel evaluation strategy is constructed to meet the national standard requirements, which focus on the "existence detection" of microcracks rather than size measurement. Experimental results on the Shenyang University of Technology Benchmark 1(SUT-B1) dataset show that the proposed method achieves an average precision of 94.69% and an F1-score of 92.60%, with only 3 missed detections and 5 false detections. Its performance not only surpasses that of baseline and mainstream lightweight models, which achieve namely, 94.53% AP and 87.85% F1-score—demonstrating the advantage of the DenseNet structure in feature reuse, but also exceeds the best performance in comparative experiments (94.56% AP and 90.90% F1-score), confirming both the advantage of DenseNet-based feature reuse and the necessity of the structural optimization strategy. The proposed approach demonstrates practicality and scalability in the field of industrial inspection, offering a new technical direction for similar fine defect recognition tasks. The related code is publicly available at: https://github.com/zhuanzhaun/Lightweight-DenseNet.

Abstract:Pulse eddy current testing, as a non-contact, environmentally friendlynon-destructive testing method requiring no coupling agent, is widely usedto assess the structural healthof metallic pipelines. The timeliness and accuracy of pulse eddy current response estimation are severely constrained by electromagnetic modelling approaches for well casing pulse eddy current testing. Traditional mathematical modelling approaches demand substantial prior knowledge, entailing complex model construction and high computational cost.Meanwhile, purely data-driven neural network methods lack physical information constraints and exhibit insufficient robustness. Field operations frequently necessitate a pulsed eddy current modelling method balancing efficiency and precision. This research addresses this issue by proposing a novel physical information neural network surrogate model. Electromagnetic physical laws are embedded as prior knowledge within the objective loss function to guidethe training process of deep neural networks. Furthermore, sub-neural networks are introducedto estimate electromagnetic responses across distinct computational domains, separated according to their physical characteristics. An interface loss function is designed to compensate for discontinuities in output between networks when predicting results across dual computational domains, thereby enhancing the accuracy and robustness of electromagnetic response estimation. The performance of the proposed physio-informative neural network was validated using electromagnetic response data obtained via finite element analysis. Its capabilities were compared against conventional purely data-driven neural networks and interpolation algorithms. Results show that the physical information neural network model accurately estimates electromagnetic responses in oil casing eddy current testing, achieving a coefficient of determination exceeding 0.95. Furthermore, the inference speed of the physical information neural network model surpasses that of finite element analysis by over 52 times.

Abstract:In promoting the healthy development of the wind energy industry, condition monitoring of wind turbines (WT) plays a crucial role. Existing data-driven condition monitoring methods primarily rely on time-series data, such as supervisory control and data acquisition (SCADA) and condition monitoring system data analysis, failing to effectively utilize the information contained in the WT textual data, such as design specifications, operation manuals, papers, patents, maintenance records, fault reports, etc. They have limitations in the analysis of fault transmission causality and the interpretability of analysis results. Therefore, this article proposes a WT condition monitoring method driven by the fusion of a knowledge graph and spatio-temporal graph neural network (KG-STGNN). This method first utilizes textual data combined with WT structural and other information to construct a WT operation and maintenance knowledge graph, forming a directed graph structure for WTs. Then, SCADA data are embedded into the graph structure to generate WT time-series graph data. A high-order graph attention network (HGAT) and a Transformer are used to establish a spatio-temporal graph, the spatial and temporal characteristics in the graph data are mined for condition monitoring. Historical healthy data of the WT are used to train the KG-STGNN model. Finally, the operating status of the WT is determined based on the information represented by nodes in the graph, and a monitoring strategy is constructed to determine fault warning times and interpret the condition monitoring results. Analysis of actual WT cases evaluates the effectiveness of the proposed method, with its performance outperforming traditional graph neural network models. Case studies on two WTs show that the proposed method exhibits excellent performance in condition monitoring, since it has the lowest false alarm rate and the earliest anomaly warning capability. Ablation experiments show that the graph structure constructed from the knowledge graph is critical for improving model performance. The proposed monitoring strategy eliminates more than 85% of false alarms and also provides good interpretability for the monitoring results.

Abstract:Accurate detection of steel surface defects is a critical aspect of industrial quality control. Especially in precision manufacturing fields such as mechanical engineering, automotive industry, electronics, aerospace, and artillery barrel production, surface quality directly determines the safety and reliability of end products. To address the limitations of existing steel surface defect detection methods, including insufficient multi-scale defect detection capability, high missed detection rates for small and low-contrast defects, and suboptimal bounding box regression accuracy, this article proposes an improved multi-scale steel surface defect detection method based on YOLOv11n. A multi-scale dynamic convolution module is designed, which employs parallel heterogeneous convolutions and a dynamic weight fusion mechanism to enhance the model′s ability to capture multi-scale defects. A dynamic residual fusion module is formulated, replacing the baseline C3K2 module with grouped convolution and a dual-residual structure. This significantly reduces the parameter count while improving multi-scale feature fusion and gradient flow efficiency, alleviating the degradation issue in deep network training. The deformable triple attention mechanism is enhanced by integrating deformable convolution and cross-dimensional interaction, enabling the attention receptive field to dynamically adjust according to defect morphology, thereby precisely focusing on small, low-contrast regions and suppressing complex background interference. The Shape-IoU loss function is adopted, which incorporates shape and scale factors to optimize bounding box regression accuracy, addressing the failure of the penalty term in traditional CIoU when the aspect ratios are identical. Experimental results on the NEU-DET dataset show that the improved model achieves an mAP@0.5 of 81.9%, representing a 6% improvement over the baseline YOLOv11n. The parameter count is only 2.3 M, and the computational cost is reduced to 5.9 GFLOPs, meeting the requirements for deployment on edge devices. Generalization experiments on the GC10-DET dataset show a 4.1% improvement over the baseline model. Visualization analysis and generalization experiments further validate its robustness and practicality in complex industrial scenarios.

Abstract:To address the challenges of complex defect morphology, strong background interference, and the strict requirements for real-time performance and lightweight deployment in printed circuit board (PCB) defect detection, an efficient lightweight detection model named KDYOLOv8 is proposed in this article, which integrates wavelet transform and knowledge distillation. Firstly, a Star-YOLO backbone network is designed, utilizing star operation to map input features into a high-dimensional non-linear space, thereby enhancing feature extraction capabilities for complex defect patterns while significantly reducing computational redundancy. Secondly, the wavelet transform convolution (WTConv) is introduced to decouple high-frequency defect details from low-frequency background textures through multi-resolution analysis, effectively suppressing noise interference and expanding the receptive field without significantly increasing parameters. Meanwhile, an EMBFPN enhanced multi-scale bi-directional feature pyramid network is constructed, employing a bi-directional information flow interaction mechanism to strengthen the fusion of deep and shallow features, addressing the problem of small defect feature dilution in deep networks. Furthermore, a channel-wise knowledge distillation (CWD) strategy is adopted to guide the lightweight model in learning the channel attention distribution of the teacher network, compensating for accuracy loss caused by model compression. Experimental results show that, on a public PCB defect dataset, KDYOLOv8 achieves a mean average precision (mAP) of 97.1%, with a model size of only 2.9 MB and an inference speed of 117.3 fps. Compared with the baseline YOLOv8n, it maintains high accuracy while reducing the volume by 52.5%. In cross-dataset generalization experiments, the detection accuracy for subtle defects such as “mouse bite” and “short” improved by 1.9% and 1.6%, respectively. This study effectively balances detection speed, accuracy, and resource consumption, providing strong support for industrial deployment in resource-constrained environments.

Abstract:Photovoltaic (PV) power generation serves as a core pillar for China′s energy transition and the achievement of the dual carbon goals, and the efficient and safe operation of its systems is of vital importance. However, PV panel faults can directly lead to a sharp drop in power generation, a rise in safety risks, and an increase in operation and maintenance costs, which have become a key bottleneck restricting the high-quality development of the PV industry. To address this issue, a PV panel fault diagnosis method based on cross-scale feature fusion is proposed. By leveraging the existing security equipment in PV power stations, this method can accurately detect six operational states of PV panels, including normal, dust accumulation, bird droppings adhesion, electrical loss, physical damage, and snow coverage, without incurring additional hardware costs. Specifically, this method firstly constructs a cross-scale feature fusion framework combining the Transformer and the convolutional neural network (CNN). The Transformer captures the global contextual features of PV panel images through its self-attention mechanism, providing global guidance for the CNN to extract local detailed features. Secondly, a dense connection mechanism is designed in the CNN branch, which enhances the propagation and reuse capabilities of features at different levels through cross-layer connections of feature maps. Meanwhile, a visible light dataset for six operating states of PV panels is developed in a targeted manner, covering different lighting conditions, weather situations, and fault types. Compared with other models, this method exhibits superior comprehensive performance, with a Top-1 accuracy of 93.33% and a Top-3 accuracy of 100%. Additionally, the model has a relatively lightweight parameter scale and low hardware computing power requirements. Furthermore, to comprehensively evaluate the engineering application value of the method, a three-dimensional comprehensive evaluation index system, model accuracy-usage efficiency-application scale, is established, which further verifies the feasibility of the method for practical deployment in PV fields.

Abstract:In pursuit of China′s Carbon Peak and Carbon Neutrality goals, solid oxide cells have emerged as key technologies for green hydrogen production and efficient power generation, due to their superior energy conversion efficiency and reversible operation capability. However, severe thermomechanical coupling during high-temperature operation causes localized Joule heating, distorted thermal fields, and accelerated degradation. Current multiphysics characterization methods for temperature and stress are constrained by the instability of sensor materials at high temperatures, leading to inadequate spatial resolution and dynamic response for accurate internal monitoring. This review systematically summarizes recent advances in temperature and stress measurement techniques for SOCs, comparing four core methods: Thermocouples are cost-effective with fast response yet prone to thermal drift; optical fiber sensing enables distributed measurement with electromagnetic immunity but suffers from limited reliability under extreme conditions; infrared thermography offers non-contact surface temperature mapping yet depends on emissivity and cannot probe internal temperatures; high-energy radiography allows high-resolution 3D stress reconstruction but requires complex instrumentation and long testing times. To address these limitations, we propose a multimodal sensing strategy integrating multiple transducers to enhance spatial resolution and thermal resilience, along with a non-destructive strain measurement approach combining high-energy radiography with digital image correlation to overcome dynamic response constraints. This work provides precise metrological support for thermal management optimization and structural reliability in SOC stacks, facilitating the development of safe, efficient, and low-carbon energy systems.

Abstract:Ultra-low expansion quartz glass (ULE) is a special material with a coefficient of thermal expansion (CTE) of 10-9/℃, achieved through precise TiO2 doping. It boasts excellent dimensional stability and temperature adaptability, making it widely used in astronomical observations, laser detection, and semiconductor manufacturing. During the preparation of ULE materials, the uniformity of TiO2 distribution affects the uniformity of CTE distribution, which in turn significantly influences the optical surface figure and the stability of imaging quality of ULE optical components. Therefore, ensuring full-bore CTE homogeneity in ULE materials has become a key focus in material preparation technology research, among which highprecision measurement of CTE and its uniformity provides strong support for efficient iterative optimization of the material preparation process. This paper provides a comprehensive review of the preparation methods of ULE and its CTE control mechanisms, focuses on analyzing the research progress on detection technology of CTE and its uniformity, and compares the advantages and disadvantages of each technology: Traditional sampling calibration methods, such as the pushing-rod method, interferometry, and photoelasticity, can achieve high-precision measurements but have limitations such as destructiveness and complex operation; while nondestructive testing methods (such as the refractive index method, the linear focused ultrasonic scanning method and the ultrasonic longitudinal wave velocity method) provide effective solutions for full-bore CTE evaluation of ULE materials. In particular, the ultrasonic longitudinal wave velocity method can effectively characterize CTE and its uniformity in large ULE boules, and achieve a good balance between measurement accuracy and engineering practicability. The paper further explores the key challenges and future prospects of ultrasonic nondestructive testing technology, proposing that future efforts should address the impact of microscopic characteristics on sound wave transmission and the challenge of achieving high-precision sound wave velocity measurements to advance the preparation and application of large-size, high-performance ULE materials. This review constructs a multi-disciplinary perspective on the detection technology system for coefficient of thermal expansion in ULE, providing new ideas and technical references for related research.

Abstract:In angle-measurement instruments,the mirror and rotary stage inevitably influence and often determine measurement accuracy.To clarify the systematic deviations introduced in the F-P etalon-based angle-measurement system,this article formulates a mathematical model describing micro-angle measurement errors caused by the mirror and the rotary stage,followed by a comprehensive analysis and correction.The model incorporates mirror pitch-angle error,initial zero-position angle error, eccentricity error,rotary-stage installation error, and rotation-axis eccentricity error.Through matrix operations and three-dimensional coordinate transformations, the displacement of the interference-ring center on the imaging plane under various errors is derived, forming a complete error-propagation framework. MATLAB is used to quantify the influence of each error source,examining cases with and without coupling.Results show that without coupling, the mirror eccentricity error dominates, whereas with coupling, principal component analysis (PCA) indicates that the mirror eccentricity error and initial zero-position angle error jointly exert the greatest impact. Based on this analysis, a targeted correction for mirror eccentricity is implemented. Within a measurement range of 1 800″, accuracy improves from ±1.93%FS to ±0.22%FS,reducing micro-angle measurement error by 91.9%. The effectiveness of the model and correction strategy is evaluated.

Abstract:Double-flank measurement is the most widely used full inspection method in gear production workshops, which can obtain the radial composite deviation of gears. However, the double-flank measurements are usually affected by the error of the master gear. Correcting the error introduced by the master gear in double-flank measurements is an effective method to improve the application value of double-flank measurement. Based on the existing double-flank rolling tester, an error correction method for the master gear in double-flank measurement, based on changing the initial measurement teeth of the master gear and continuous rotation measurement was proposed. The corrected values of master gear error that can be applied to double-flank measurements are obtained. Multiple sets of simulation experiments with different radial composite deviations for the master gear and product gear are designed. The influence of radial composite deviation on double-flank measurements and the mechanism of the master gear error correction method are revealed. The consistency evaluation parameter that can reflect the accuracy of the double-flank measurement system is proposed to evauate the effectiveness of the error correction method. The measurement experiments show that, after using this method to correct double-flank measurements, the consistency of the center distance curve cluster measured from different phases of the master gear and the same phase of the product gear is significantly enhanced. Compared with before error correction, the consistency evaluation parameter shows that the accuracy of the double-flank measurement system has increased by 77.82%; compared with existing error compensation methods, the consistency evaluation parameter of the center distance curve cluster has been improved by 52.08%. The error correction method for master gear in double-flank measurement has effectively reduced the impact of the master gear error on the double-flank measurements. It has enhanced the traceability of double-flank measurements and the application value of double-flank measurement method, improved the accuracy of the double-flank measurement system, and enabled the double-flank measurements to more accurately characterize the quality of the product gear. Furthermore, this method exhibits superior performance in terms of measurement efficiency and the effectiveness of error correction.

Abstract:In mechanical systems, assembly errors are one of the key factors affecting the operational accuracy and overall performance of the entire machine. However, at present, in the prediction of assembly errors, there is a problem that the predicted values differ greatly from the actual values, which seriously affects the assembly quality and the subsequent performance optimization. Aiming at this problem, an assembly accuracy prediction method based an improved RPM-Net and multi-constraint assembly surface weight distribution is proposed. In terms of obtaining the fit error of parts, the point cloud registration technology is adopted to precisely align the point cloud of the assembly surface, so as to obtain the geometric relationship and error information between the assembly surfaces. During the registration process, by embedding the attention mechanism into the RPM-Net framework, the network′s focus on important feature information is enhanced. This effectively suppressees the risk of registration falling into local optimum, significantly improves the registration accuracy of the assembly point cloud, and provides a more reliable data basis for the subsequent error transmission analysis. In terms of the error transmission of part fit, based on the influence degree of the end pose of each assembly surface, the torsors of the assembly surfaces that constitute multiple constraints are weighted and summed, and the torsors of non-repetitive constraints are compounded. This processing method enables comprehensive consideration of the torsors of each assembly surface, avoiding the prediction deviation caused by ignoring certain assembly surface torsors in traditional methods, and thereby improving the accuracy of assembly precision prediction. The experimental results show that the deviations between the predicted and actual values in the five degrees of freedom in space are precisely bounded within 2 μm and 4×10-5 rad. Compared with the traditional series method and algebraic operation method, the proposed method more accurately reflects actual assembly situation, providing an effective model reference for the subsequent improvement of assembly accuracy.

Abstract:As the core component of soft robots, the deformation and motion characteristics of soft actuators directly determine the operational performance of the robots. To overcome the high driving pressure and limited structural design flexibility in traditional pneumatic artificial muscles, this study proposes a rectangular bellows actuator. Through finite element simulation comparing the performance of triangular, parabolic, rectangular, and semicircular bellows structures, the results indicate that the rectangular structure achieves the optimal comprehensive evaluation coefficient, combining excellent axial elongation capability with radial stability. Experimental verification shows that under pure pneumatic loading, the actuator can reach a maximum elongation stroke of 53.2% at 60 kPa, with the elongation exhibiting a linear relationship with the number of bellows and a nonlinear relationship with the input pressure. Further simulation analysis reveals that under the coupled action of air pressure and tangential load at the free end, the actuator can effectively suppress bending deformation by adjusting the internal pressure. An equivalent bending stiffness model is subsequently established, indicating that the stiffness decreases with increasing bellows number while exhibiting an approximately linear increase with rising internal pressure. Accordingly, an equivalent bending stiffness model is established, which demonstrates that the equivalent bending stiffness decreases with an increase in the number of bellows and shows an approximately linear increasing trend with the rise in input pressure. In terms of dynamic modeling, a motion equation for the actuator under compressive load at the free end is constructed based on a three-element model, and the model parameters are identified using the recursive least squares method. Experimental results confirm that the proposed accurately predicts the dynamic response of the actuator under different load conditions, with a displacement prediction error of approximately 2 mm. Through the integration of structural design, simulation analysis, and experimental validation, the static and dynamic deformation laws of the rectangular bellows actuator are systematically elucidated, and corresponding equivalent bending stiffness and three-element dynamic models are established, laying a theoretical foundation for its further application in driving and execution fields of soft robotics.

Abstract:Continuum robots suffer from low control accuracy and poor safety performance due to large modeling errors, complex and variable shapes, and susceptibility to external dynamic disturbances, which makes it challenging to achieve precise operations and safe interactions in confined or complex environments. To address these issues, a self-sensing approach for continuum robot shape estimation based on IMU measurements and a piecewise polynomial curvature model is proposed, enabling the detection and reconstruction of their three-dimensional curved shapes. First, a shape detection system is designed by deploying multiple IMUs along the continuum robot, and the PPC model is employed for kinematic modeling and analysis to accurately characterize non-uniform bending deformations. To estimate the robot′s bending profile and end-effector position, a self-sensing approach for shape estimation that fuses IMU measurements with a PPC model is introduced. In this framework, the modal coefficients of each curvature segment are determined from a limited number of attitude observations, thereby reconstructing the overall shape of the robot. Finally, an experimental platform for shape detection is established, and the proposed method is validated through theoretical simulations and multiple experiment trials. The results demonstrate that the proposed approach achieves reliable performance under various bending angles and loading conditions, with an average reconstruction error of approximately 2.5 mm and a deviation below 3 mm under loaded scenarios. In addition, dynamic bending experiments further validated the peoposed method′s real-time capability and shape tracking performance during continuous motion, with an average end-effector position error of approximately 2.57 mm. This validates the effectiveness and accuracy of the constructed motion model and proposed shape detection method, providing a reliable shape sensing foundation for precise operation and closed-loop control of continuum robots in constrained environments.

Abstract:Addressing the issues of single-dimensional environmental information perception, inefficient dense map creation, and insufficient environmental detail information in sparse maps during visual perception of external environmental changes for legged robots in outdoor environments, this paper proposes a machine vision-based multi-dimensional environmental mapping method for legged robots that achieves global large-scale and local high-resolution capabilities. The method employs a visual SLAM algorithm combined with RGB images and depth information to achieve camera pose estimation and generate environmental point clouds. Furthermore, by employing the improved voxel filtering to reduce point cloud density and utilizing ray projection to create virtual points, the method realizes global large-scale and local high-resolution environmental geometric dimension map creation. Based on this foundation, targeting the requirements for environmental physical dimension information perception of outdoor legged robots, the method implements high-precision semantic segmentation of outdoor terrain environments through an improved SegNet network. It further utilizes terrain optical characteristics and surface structural features to establish a mapping from terrain semantics to terrain physical layer attribute parameters through a decision model, thereby achieving the creation of terrain physical dimension maps. Finally, through the fusion of terrain geometric dimension maps and physical dimension maps, the creation of multi-dimensional environmental map for outdoor legged robots is accomplished. The rationality and effectiveness of the proposed mapping method are validated through physical platform mapping experiments. The experimental results demonstrate that the proposed multi-dimensional environmental mapping method exhibits significant advantages over traditional mapping methods in terms of mapping performance, environmental key information extraction, and perception dimensions. It is more suitable for improving legged robots′ comprehensive non-contact understanding of environmental information during outdoor movement, thereby enhancing the environmental adaptability of legged robots in outdoor environments.

Abstract:This paper proposes a multi-source fusion SLAM method based on LiDAR/IMU/UWB tight coupling to address the problem of decreased positioning accuracy caused by dynamic interference in urban environments. The aim is to improve the robustness and localization accuracy of the system under dynamic interference. In the odometery stage, this paper constructs a factor graph framework that deeply integrates IMU pre integration and multi base station UWB distance observation. IMU provides high-frequency motion priors to compensate for point cloud distortion, while UWB introduces stable external constraints through absolute ranging without cumulative errors, which can effectively suppress the cumulative errors of LiDAR in dynamic environments and significantly improve the robustness of pose estimation in dynamic interference scenarios. In addition, in the loop detection and global optimization stage, this paper designs an extended descriptor that combines UWB information to encode and fuse point cloud geometric features with UWB absolute range and signal strength information, forming a more discriminative scene representation. Based on this descriptor, a two-level loop detection strategy of "coarse retrieval fine verification″ is adopted: First, UWB information is used to quickly screen candidate loop frames, and then point cloud descriptor geometry verification is performed, effectively improving the accuracy of loop detection in dynamic and homogeneous environments. After detecting the loop, the UWB historical information and point cloud constraints are jointly incorporated into the global factor map for collaborative optimization, further improving trajectory consistency and map closure accuracy. Experiments real-word urban dynamic-scene datastes show that our method can significantly reduce absolute trajectory error (ATE) compared with methods such as FAST-LIO2 and MR-ULINS. Meanwhile, the UWB assisted loop back detection mechanism performs better on the precision recall curve, effectively reducing dynamic ghosting and false matching, verifying the effectiveness and superiority of this method in urban dynamic environments.

Abstract:The aerial manipulator is a crucial equipment for dismantling and installing current-carrying clip bolts during substation maintenance operations. To address the issues of slow error convergence and performance degradation caused by unknown disturbances during the visual servoing tracking of the bolts, this paper proposes a non-singular terminal integral sliding mode control method based on an integral sliding mode disturbance observer. The proposed method establishes the kinematic model of the image-based visual servoing system and analyzes the effect of external disturbances on model uncertainty, which is further characterized using an equivalent disturbance term. A non-singular terminal integral sliding mode controller incorporating an integral power term and a non-singular terminal term is designed. Combining with an exponential reaching law, the non-singular terminal integral sliding mode control law is derived, guaranteeing the rapid convergence of the system tracking error in finite time. Furthermore, an integral sliding mode variable is used to construct a dynamic observation equation, from which the system state and disturbance estimation equations of the integral sliding mode disturbance observer are derived. The estimated disturbance values are then feedforward compensated into the sliding mode control law, enhancing the system′s disturbance rejection performance. The stability of the system and the finite-time convergence of the tracking error are proven using Lyapunov theory. Finally, simulation experiments verify the feasibility of the proposed method. Additionally, an aerial operation simulation platform is constructed to design dynamic target tracking experiments under various disturbance conditions. The experimental results show that the proposed method reduces the average convergence time by 4.92 s and decreases the root mean square error by an average of 11.03 pixels compared to the benchmark methods. It consistently improves convergence speed and tracking accuracy under various unknown disturbance conditions. Furthermore, the designed sliding surface exhibits favorable dynamic performance, and the disturbance observer accurately estimates external disturbances, indicating strong potential for practical engineering applications.

Abstract:In urban environments where satellite signals are limited, road network maps serve as structurally stable prior information that effectively suppresses the cumulative drift of visual odometry. However, existing road-network-assisted localization methods often suffer from imprecise turn detection and simplified topology-matching strategies, making them difficult to adapt to complex road structures. To address these issues, this paper proposes a road-network-assisted vehicle visual localization method integrating heading awareness and topological matching. A sliding-window strategy combined with heading-rate variation is employed to determine the vehicle motion state, while a straight-curve ratio is introduced to characterize the geometric properties of turning segments. Representative turning points are extracted using the maximum deviation measure, providing reliable structural cues for subsequent node association. Based on this, a topology matching method that accounts for both geometric consistency and structural similarity is developed. A perpendicular nearest-point constraint is applied to achieve accurate position association in straight-driving scenarios, while a multi-level topology similarity metric ensures robust structural matching during turns, significantly enhancing consistency in complex intersections, gentle curves, and high-curvature segments. Furthermore, a Kalman filter is employed to fuse short-term visual odometry observations with global structural constraints from the road network. Road-segment orientation information is utilized to suppress heading drift during straight driving, resulting in a robust localization framework that tightly couples heading perception and topology matching. Experiments conducted on the KITTI dataset and campus field tests demonstrate that the proposed method effectively suppresses the accumulated drift of visual odometry while maintaining real-time performance. The localization accuracy improves by 78.7% compared with raw visual odometry and by 38.3% and 34.0% compared with the MPF and RNAP road-network-assisted methods, respectively. Stable performance is preserved even under node deviations or partial road-network loss, confirming the reliability and general applicability of the proposed method in various complex road conditions.

Abstract:To address the time-energy-smoothness path optimization and trajectory motion control challenges in mobile robot navigation under uncertain conditions, this study constructs map environments and available paths using Delaunay triangulation. A free-space judgment criterion is proposed, and a collision-free model is established. The paper presents optimization methods that include acute vertex deletion, path replacement, and redundant point removal, as well as a cubic NURBS path fitting approach. Furthermore, non-probabilistic reliability is introduced to evaluate path states, with optimal reliability paths and their weighting concepts defined and explained. A cost function integrating path task time risk and energy risk metrics is designed. Meanwhile, constraints such as peak curvature limits, restrictions on curvature change rates, and safety distances are integrated into the model. The optimal path smoothing planning and a five-stage S-type acceleration-deceleration motion control with a jerk that satisfies the ′Bang-Bang-Singular′ strategy are carried out. Experimental results demonstrate that our method achieves a 1.907% reduction in time risk and a 40.57% decrease in energy risk compared to approaches employing quadratic B-spline path planning with acceleration control that satisfies the ′Bang-Bang-Singular′ strategy, and the movement is smoother, safer, and more efficient. In contrast to trajectory planning using the VGSP algorithm, the time risk index shows a slight increase, while the energy risk index decreases by 86.46%, with improved safety guarantees for robot operation. Field tests further validate the effectiveness of our method in solving optimal trajectory planning for mobile robots under stringent constraints, ensuring both smooth trajectory curves and the system′s dynamic and flexible performance during motion. This approach achieves unified geometric-motion planning and optimal non-probabilistic reliability regarding task-time and energy.

Abstract:Fuel assemblies, serving as the heat-releasing components within nuclear reactors, constitute the energy source for nuclear power generation. Among these, the zirconium-based cladding of nuclear fuel acts as the primary safety barrier in nuclear power plants. It effectively prevents the dispersion of fission products while conducting heat away, thereby avoiding fuel corrosion due to cooling. Therefore, the structural characteristics of cladding tubes are closely related to the performance of fuel assemblies. Conducting effective non-destructive testing on zirconium-based fuel cladding to achieve high-precision wall thickness and electrical conductivity characterization remains a critical challenge in nuclear safety inspection. This article investigates a novel stacked eddy current sensor for the conductivity and thickness measurement of a cladding tube. Firstly of all, a stacked array eddy current sensor consisting of three absolute coils is designed. Based on the structural configuration of this sensor, the analytical solution is established for cladding tube detection. Based on the analytical solution, the s cross-frequency of self-inductance can be extracted in sweep mode. Moreover, the logarithm of the cross-frequency exhibits a linear relationship with the lift-off distance. The slope of this linear relationship depends solely on the wall thickness and is independent of the conductivity of the material. Consequently, the wall thickness can be estimated by the slope of the fitting line. Subsequently, the eddy current testing problem can be transformed into a least-squares problem for the cladding tube. Taking the wall thickness measurement results as prior information, the conductivity of the cladding tube is inverted through the improved Newton iteration algorithm. Finally, the eddy current testing experiment platform is established to validate the effectiveness of the proposed method, and the results show that the maximum measurement error of the proposed method is only 1.3%.

Abstract:Wireless power transfer (WPT) system with quadrature double channels (QDC) features a controllable magnetic field distribution, enabling a distribution of strong coupling regions according to the locations of the receiver coils, and thereby supporting wide-range power delivery for electrical devices. Circulating currents are induced by the aligned, cross, and same-side mutual inductances within the QDC coupling structure. The resonance state and transmission efficiency in quadrature double channels-Wireless power transfer (QDC-WPT) system are deteriorated by the circulating currents, unbalancing the power between the two transmission channels. This article formulates the QDC-WPT model and analyzes the component and formation mechanisms of circulating current under two excitations. The reactive and active components of the circulating currents are clarified. Expressions for the power transmission component and six sets of circulating current are derived. Accordingly, a configuration method for resonant circuits is proposed to compensate for the reactive current. The distribution characteristics of active circulating currents with respect to the relative positions are analyzed, revealing three mechanisms by which the circulating currents act upon the power transfer component: dominant, superimposed, and cancelled out. Then, the zero-power point under the circulating currents for the single energy channel is derived, with the power balance constraint condition for the dual energy channel. The QDC-WPT system operating in power transfer mode, critical mode, and power feedback mode, are determined, respectively. Finally, a 1 kW system model and a prototype are established. Simulation and experiment results validate the proposed circulating current mechanism, compensation method and system performance analysis.

Abstract:Millimeter wave radar has become a research hotspot in vital sign monitoring due to its advantages of non-contact operation, high penetration, high precision, and real-time capability. It has already been applied in medical cardiovascular health assessment, tumor detection and localization, and health monitoring for newborns and children. However, during indoor monitoring, the millimeter wave radar echo signals are susceptible to subtle human body movements and multipath effects, resulting in various noise components that complicate the extraction of heartbeat information during sleep. This paper proposes a vital sign detection method for millimeter wave radar based on Empirical Wavelet Transform and Variational Mode Decomposition (EWT-VMD) for detecting respiratory and heartbeat signals from the echo signals. By establishing a vital sign echo signal model for millimeter wave radar detection, the frequency composition of the signal is analyzed. Experiments are conducted using a 77 GHz millimeter-wave radar to acquire echo signals, and frequency-domain coherent accumulation is applied to two types of chirp signals within the same signal frame, which simultaneously suppresses noise interference and enhances the amplitude of useful signals. The empirical wavelet transform (EWT) is employed to decompose the vital sign-containing signal, followed by signal reconstruction after clutter elimination. The wavelet-reconstructed signal is further processed by variational mode decomposition (VMD) to extract the human respiratory rate (RR) and heart rate (HR) respectively. The extracted results are compared with the signals collected by an electrocardiogram (ECG) to assess the extraction accuracy. Experimental results across various scenarios demonstrate that the proposed frequency-domain coherent accumulation method effectively enhances echo signal SNR. The combined CWT-VMD algorithm successfully detects RR and HR with a detection accuracy reaching 94%.

Abstract:Addressing the challenge of achieving high-precision tracking for low-altitude unmanned aerial vehicle (UAV) trajectories in complex environments, this paper proposes a residual-driven adaptive Kalman filter(RD-AKF) method based on a distributed microphone array. Within this framework, acoustic source observations are acquired through two complementary algorithms of different dimensions: a geometry-based triangulation method that provides directional constraints, and a frequency-domain phase compensation beamforming(FD-PCB) algorithm that captures energy spatial distribution characteristics. These physically complementary observations are integrated as joint measurements into a unified Kalman filtering framework. To mitigate the limitations of fixed-parameter models in handling abrupt noise from dynamic targets and to improve tracking accuracy, the method incorporates a residual-driven adaptive mechanism into the Kalman filter. This mechanism computes the prediction residuals of each observation channel in real-time and dynamically adjusts the measurement noise covariance matrix based on their statistical properties, thereby optimizing data weighting for the joint estimation of the acoustic source′s position and velocity. Comparative experiments were conducted in an indoor environment by tracking the same low-altitude UAV flight trajectory, comparing the proposed method against standalone baseline intersection, FD-PCB localization, and standard Kalman filtering with fixed covariance matrices. The results show that the Kalman-fused approach reduces the root mean square error(RMSE) by at least 24.3% compared to individual algorithms. With the residual-driven adaptation, the RD-AKF further reduces the RMSE by 18.5% and the maximum positioning error by 15.6% compared to the standard Kalman filter, achieving significantly improved tracking accuracy and stability while maintaining reasonable computational cost. The proposed method provides a high-precision and robust solution for dynamic acoustic source tracking in complex scenarios.

Abstract:Fiber Fabry-Perot pressure sensors exhibit high sensitivity to dynamic pressure and can accurately capture instantaneous pressure variations; thus, they are widely used in dynamic pressure measurements. Based on this capability, the present study employs an optical fiber Fabry-Perot pressure sensor to acquire and measure dynamic impact pressure. To facilitate field testing of the impact pressure and frequency response of the fiber Fabry-Perot pressure sensor, the proposed method establishes an instantaneous frequency-swept interference model and analyzes the time-varying characteristics of the interference spectrum induced by dynamic variations in the cavity length. By performing a Fourier transform on the raw interference spectrum, applying a Hamming window to retain the positive-frequency components, and subsequently executing an inverse Fourier transform to obtain the analytic signal and its phase, the dynamic cavity length can be extracted. To further obtain the sensor′s frequency-response characteristics, a Taylor expansion of the interference spectrum is conducted, and frequency components obtained from the Fourier transform are processed through addition and subtraction operations to derive the sensor′s frequency response. The entire demodulation procedure requires no Doppler error compensation; instead, the dynamic cavity length and frequency response can be rapidly retrieved solely through Fourier-transform operations, enabling a maximum demodulation rate of up to 10 MHz. Finally, the proposed demodulation model is compared with a single-wavelength intensity demodulation method. The relative frequency-demodulation errors of the two pressure sensors are less than 1% and 0.7%, respectively, while the relative pressure-demodulation errors are below 0.1% and 0.08%. Both frequency and pressure demodulation exhibit very small errors. Experimental results demonstrate that the proposed demodulation model is highly suitable for rapid measurement of impact pressure and frequency response by using fiber Fabry-Perot pressure sensors. Compared with conventional demodulation approaches, the proposed model achieves higher speed and a simpler operational procedure, making it more appropriate for field environments with harsh conditions.

Abstract:In recent years, fiber Bragg grating (FBG) sensors have been widely used in structural health monitoring (SHM) of aerospace structures due to their advantages, such as compact structure, electromagnetic interference resistance, and quasi-distributed integration. However, when exposed to harsh environments for a long time, they are susceptible to temperature, vibration, and other factors, resulting in problems such as spectral noise and baseline drift, which reduce the signal-to-noise ratio (SNR) and seriously affect demodulation accuracy. Traditional denoising algorithms, such as Savitzky-Golay filter and wavelet transform, rely on manual parameter setting in low-SNR scenarios, resulting in poor adaptability and difficulty in meeting high-precision monitoring requirements. This article proposes a novel DSC-U-Net deep neural network model, which integrates the feature extraction capability of the U-Net architecture and the lightweight advantage of depthwise separable convolution (DSC), enabling effective removal of noise and baseline distortion. Based on the coupled-mode theory and transfer matrix method, 90 000 spectral samples covering an SNR range of -20 dB to 20 dB are simulated for model training and testing. Subsequently, the data volume required for model training and the model training results were discussed. Test results show that the model trained with the full dataset can improve the SNR of 0 dB spectra to 13.266 dB, with a similarity of 0.892 to pure spectra and a root mean square error (RMSE) of only 0.05, outperforming traditional algorithms such as arPLS combined with window functions. An experimental system for harsh environments (-55℃~150℃) is established for experimental data collection, and verification is conducted using the model trained with simulated data. The demodulation algorithm combining DSC-U-Net and MLP reduces the demodulation error from 0.297 pm to 0.023 pm, with an accuracy improvement of 92.26%. Model training using simulated data can significantly reduce training costs. The DSC-U-Net deep neural network model requires no manual intervention, featuring both high precision and efficient computation. It solves the demodulation problem of FBG signals under low SNR and provides a reliable solution for long-term stable monitoring in harsh aerospace environments.

Abstract:To address the degradation and failure of pressure sensors in the sealing water system of dredger mud pumps, an interpretable fault diagnosis and health assessment framework that integrates hydraulic physical principles of sealing pumps with data-driven modeling is established. Firstly, a physical model of sealing water pressure is constructed based on the hydraulic characteristics of the sealing water pump to derive the sealing water pressure at the pump outlet, which serves as a physically groundedreference reference independent of the sensor health state. Considering the nonlinear degradation and parameter drift of the equipment during long-term operation, dynamically optimized power-law fitting coefficients are further introduced to improve the adaptability and accuracy of the physical model under different operating conditions. A multi-scale convolution-Transformer attention fusion network incorporating physical information constraints is developed, where a piecewise dynamic loss-weight scheduling strategy is introduced to achieve synergistic optimization between physical consistency and data-driven fitting accuracy, thereby significantly enhancing sealing-water-pressure prediction accuracy and improving generalization capability under complex operating conditions. Based on the residual variations among the theoretical pressure, the model-predicted pressure, and the sensor-measured pressure, a threshold-based discrimination mechanism is established to realize explainable identification of transient, intermittent, and permanent faults. Meanwhile, a multi-parameter fusion reliability calculation method is proposed to quantify the full-life degradation trajectory of the pressure sensor from a healthy state to initial abnormality, accelerated deterioration, and functional failure, from healthy state, initial abnormality, accelerated deterioration to functional failure and the reliability curve can accurately present the degradation evolution characteristic of "slow-accelerated-sharp decline″. Simulation results based on real vessel operation data show that the proposed method outperforms the comparison models in prediction accuracy, stability, and convergence, with R2 up to 0.952 7, and can identify short-term anomalies in the healthy stage, realizing high-confidence fault diagnosis and intelligent maintenance support for pressure sensors.

Abstract:To overcome at the limitations of traditional electronic monitoring methods in geophysical observations, this study develops a fiber optic inclinometer system based on a Fabry-P-rot (F-P) interferometric cavity. The system combines a vertical pendulum mechanism with dual F-P interferometric cavities, using differential measurement techniques to effectively suppress environmental noise and temperature drift. The interferometric optical path and phase demodulation algorithm are optimized to significantly enhance the precision of tilt angle detection. Experimental results demonstrated that the system achieves a tilt sensitivity of 2 435 nm/″ and a resolution of 0.000 078 ″ within a measurement range of -181.6″ to 181.6″, showing excellent precision and stability. Continuous monitoring experiments conducted at the Tai′an Seismic Monitoring Center Station successfully recorded periodic solid earth tide variations, with the waveforms closely matching those observed by a VP-type electronic inclinometer, confirming the system′s high resolution and low drift characteristics. Furthermore, multiple seismic events were captured during the experiment, demonstrating the feasibility of simultaneous monitoring of solid earth tides and seismic activity. In comparison with the traditional VP-type electronic inclinometer, the fiber optic inclinometer not only accurately recorded standard seismic signals but also detected small seismic signals, indicating a clear advantage in terms of sensitivity and noise suppression. This fiber optic inclinometer system provides a new tool for crustal deformation measurement and seismic observation, with high resolution and sensitivity that meet seismic industry standards. The experimental results show that the system can effectively monitor solid earth tide waveforms and seismic events, demonstrating significant potential for widespread application. This technology provides a reliable basis for earthquake monitoring and crustal dynamics research and is expected to contribute to innovations in future geophysical observation equipment.

Abstract:Using floating-ground voltage sensors to measure the voltage of distribution network transmission lines greatly reduces cost and size compared with traditional voltage transformers. However, their capacitive voltage division principle makes measurement results susceptible to environmental interference, especially in rainy conditions. The power frequency measurement results of the sensor show rapid and significant fluctuations, which are easily confused with actual grid faults and pose challenges for distribution network operation monitoring. To address this issue, based on the sensor′s voltage division model and via theoretical analysis and simulation, this paper firstly based on the sensor′s voltage division model and via theoretical analysis and simulation, identifies that severe distortion of the sensor′s measured electric field in rainy conditions and interference to the ground-coupled capacitance of the sensor′s upper and lower plates are the main causes of the sensor′s measurement errors. Secondly, through analyzing the voltage division model and sensor transfer function, it designs a new-type voltage sensor with an equipotential shielding case and displacement current transimpedance amplification is designed. The equipotential shielding case reduces the impact of sudden external environmental changes on coupled capacitance, while the transimpedance amplification circuit and displacement current signal processing enhance anti-interference capability. Finally, an experimental platform is established, including an adjustable simulated rain device, a standard sine-wave voltage source, and a data acquisition system to test the new sensor and verify the rationality of its design scheme. Using standard 5.77 kV power frequency input voltage, comparative experiments under different rainfall intensities and simulated rain positions validate that the new sensor can effectively suppress rain interference, with measurement fluctuations reduced by over 50%. It also has good measurement accuracy, stability, and linearity, with the measurement error of the 5.77 kV power frequency voltage amplitude within 3.3%.

Abstract:In application scenarios that require stop-on-target behavior and controllable micro-displacements, ultrasonic micromotors (USM) typically achieve high-precision positioning through stepping operation. However, during actual stepping, deformation of the stator surface introduces detection errors in the capacitive angular sensor, thereby limiting the positioning accuracy. To address this issue, this paper focuses on an ultrasonic micromotor with specialized structural features and proposes a stepping control method that explicitly accounts for the initial stator deformation error. First, based on the sensor structure and the principle of a parallel-plate capacitor, the formation mechanism of angular detection errors during the start-stop phases is analyzed. The variation of the rotor-stator gap induced by the excitation and attenuation of the traveling wave in these phases is identified as the main source of error. Accordingly, an error-compensation scheme combining pre-excitation compensation and braking-parameter calibration is proposed. Second, for this type of ultrasonic micromotor, an input-output mapping is established between the excitation voltage frequency, amplitude and phase and the rotational speed, and the voltage amplitude is selected as the control variable. In the closed-loop stepping strategy, taking into account the actual output characteristics of the capacitive angular sensor, a commutation-free, smoothly transitioned stepping scheme is designed. The peaks and troughs of the capacitive waveform are selected as reference points for angle detection, and a nonlinear proportional integral derivative(PID) controller with a tracking differentiator is introduced to realize unidirectional forward trajectory planning, while adaptive adjustment of the voltage amplitude is employed to achieve precise tracking of the angular position. Finally, an experimental platform incorporating a high-precision photoelectric autocollimator is constructed. Experimental results show that the pre-excitation compensation reduces the amplitude of capacitance fluctuations by approximately 82%; in 30 trials of 12° stepping experiments, the control error remains below 0.2°, representing an improvement of about 33% compared with open-loop control, and no error accumulation is observed.

Abstract:The bearingless flux-switching permanent magnet motor (BFSPMM) places the permanent magnets on the stator side of the motor, providing an excellent heat dissipation path and fundamentally solving the problem of demagnetization of the rotor permanent magnets at high speeds, ensuring reliability under high-temperature conditions. Meanwhile, it combines the high power density and high torque characteristics of flux-switching motors with the non-contact and wear-free advantages of bearingless technology, theoretically enabling multi-degree-of-freedom active suspension of the rotor. As the core of the next-generation high-speed electric spindle, it will directly enhance the processing accuracy and efficiency of high-end CNC machine tools. In flywheel energy storage systems, it can achieve ultra-high-speed operation, significantly increasing energy storage density. Additionally, in special fields such as aerospace, semiconductor manufacturing, and life sciences, which have strict requirements for pollution-free and ultra-high-speed operation, it has irreplaceable potential and is one of the key technologies driving the upgrade of future high-end equipment. However, conventional BFSPMMs can only achieve two-degrees-of-freedom (2-DOF) magnetic levitation for the rotor. To address this limitation, this article constructs a novel Bearingless Flux-Switching Permanent Magnet Motor with Five Degrees of Freedom (5-DOF-BFSPMM) by employing two BFSPMMs and one axial magnetic bearing, and proposes a high-performance cooperative levitation control strategy for five-degree-of-freedom control. Two BFSPMMs are coaxially arranged and installed with a 3° circumferential offset to cancel their cogging torques mutually. Based on the unique structure of the motor, the 5-DOF dynamic model of the 5-DOF-BFSPMM rotor is derived. Furthermore, leveraging this dynamic model, a cooperative levitation control strategy is proposed, which decouples the rotor′s translational displacements and tilting angles. Experimental results show that the proposed cooperative levitation control strategy successfully achieves high-performance magnetic levitation control for all five degrees of freedom in the 5-DOF-BFSPMM. Specifically, the steady-state radial displacement ripple is reduced by 41.9%, and the radial disturbance amplitude is suppressed by 32.9%.

Abstract:Aiming at the issue of significant backflow power in dual-active-bridge converters under advanced dual-phase-shift control when input voltage, output voltage, and transformer ratio are mismatched, a backflow-power optimization control strategy is proposed. First, new dual-phase-shift ratios are redefined based on the phase relationship between the midpoint output voltages of the primary and secondary H-bridges and the internal phase-shift angles of the full bridges. The operating modes are divided into eight intervals. Then, four intervals for forward power transfer are selected for analysis, deriving mathematical expressions for transmission power, backflow power, and current stress. Subsequently, a segmented optimization approach is employed to determine the optimal phase-shift angle combinations in different intervals, yielding the expression for minimum backflow power. An advanced dual-phase-shift modulation strategy with minimum-backflow-power optimization is designed, enabling adaptive selection of the optimal operating interval and corresponding phase-shift angles. Finally, an experimental prototype of the dual-active-bridge converter is built. Results show that compared to traditional dual-phase-shift control, the proposed strategy significantly reduces current stress and backflow power while improving efficiency. At a voltage conversion ratio of 2.5, current stress is reduced by 48.3%, backflow power by 89.1%, and efficiency is improved by 9.4% in the low-power range, in the medium-power range, current stress decreases by 30.3%, backflow power by 92.5%, and efficiency improves by 10.7%. At a ratio of 1.5, current stress is reduced by 34.8%, backflow power is completely eliminated in the low-power range, and efficiency rises by 8.1% in low-power range, in medium-power range, current stress reduces by 45.3%, backflow power by 92.9%, and efficiency increases by 9.3%. The results validate the correctness and effectiveness of the proposed design.

Abstract:Motion error signals of CNC machine tools exhibit strong nonlinear and nonstationary characteristics, with scarce labeled data. Conventional methods struggle to effectively separate and accurately identify error sources. To address this limitation, a motion error traceability model for CNC machine tools is proposed based on chaotic attractors and transfer learning. First, one-dimensional time-series signals of circular motion errors are mapped into a reconstructed phase space to obtain chaotic attractor phase portraits. Structural features of chaotic attractors, which characterize the intrinsic dynamics of different error sources, are extracted to establish strong correlations with potential error mechanisms. These features form the basis for error source identification. Then, to address the issue of low traceability accuracy caused by overlapping chaotic attractors and significant scale variations among motion errors, a deep learning identification model based on an improved Faster R-CNN is constructed. ResNet50 and a feature pyramid network are integrated to enhance the recognition capability of chaotic attractors. Finally, to overcome the scarcity of labeled samples in CNC machine tools, transfer learning is introduced. Pre-training on the COCO2017 source domain and freezing of the shallow layers enabled effective knowledge transfer to the target domain of attractor phase portraits for motion error classification. This strategy significantly improves traceability performance under limited data conditions. At an intersection over union (IoU) threshold of 0.5, the proposed model achieves average precisions of 98.80%, 99.64%, 97.58%, and 99.97% for four typical motion error types: servo mismatch, reverse spikes, backlash, and cyclic error. Experimental analysis shows that motion errors of CNC machine tools can be effectively represented by chaotic attractors. The proposed model achieves high error-source identification accuracy under various error conditions and demonstrates strong robustness.The model exhibits high error source identification accuracy under various error conditions with strong robustness.