To address the high computational complexity of monocular depth estimation models and their difficulty in real-time deployment on resource-constrained devices， a lightweight self-supervised monocular depth estimation method based on partial convolution and multi-scale feature fusion is proposed. A self-supervised training framework is constructed by jointly optimizing a depth estimation network and a pose estimation network， where photometric reprojection error is used as the supervision signal for model training. A multi-branch feature extraction module named Lighting Block is introduced into the depth estimation network， which integrates a partial convolution branch and a dilated convolution branch to collaboratively capture local structural details and global contextual information. Experimental validation is conducted on the KITTI dataset. The results show that the proposed method achieves an Abs Rel of 0.093 and an RMSE of 4.052， with only 3.3 million parameters and an inference time of 23.6 ms. The method achieves a favorable balance between prediction accuracy and computational efficiency， demonstrating potential for applications in augmented reality， robotic vision， and industrial 3D perception.

To address the problems of scattered empirical knowledge， low fault localization efficiency， and the difficulties in reflecting multi-system coupling states through single-parameter health evaluation in fault diagnosis for slag cleaning devices， an intelligent diagnosis method integrating fault tree analysis and a dynamic health degree model is proposed. First， a fault tree model is established based on fault mechanisms， and a fault knowledge base is constructed using minimal cut sets and importance analysis. The causal relationships in the fault tree are then transformed into generative rules， and a diagnostic strategy combining forward reasoning with operation-stage identification is designed. Next， a dynamic health degree evaluation model is developed to perform health degree mapping for analog and digital parameters， incorporating coupling corrections and joint fault penalty mechanisms to achieve dynamic quantitative assessment of equipment health status. Finally， the effectiveness of the proposed method is validated through an intelligent diagnosis platform and typical scenarios. The results show that the proposed method can effectively diagnose typical faults of the slag cleaning device， improve diagnostic efficiency， and better characterize the equipment health status under coupled abnormal conditions.

Sample sets constructed by traditional multi-scale feature extraction techniques suffer from information redundancy， and some features fail to effectively characterize the bearing degradation process， which severely restricts the performance of remaining useful life prediction models. To address these issues， a bearing life prediction method combining a multi-scale feature selection strategy and the xLSTM model is proposed. First， effective features with strong degradation representation capability are selected from the original multi-scale feature set based on monotonicity criterion and temporal correlation analysis. Second， the filtered effective feature set is fed into the xLSTM prediction model， and the nonlinear mapping relationship between features and remaining useful life is explored via the deep learning capability of the model， thereby achieving accurate prediction of bearing life. Finally， the effectiveness and generalizability of the proposed method are validated. Comparisons with LSTM and Transformer-LSTM models under various operating conditions demonstrate that the proposed method achieves higher prediction accuracy.

This paper studies the security of defect classification models under adversarial attacks. Aiming at the vulnerability of industrial visual classification models to adversarial perturbations which can lead to significant declines in recognition performance， this study selects some typical attack methods—Fast Gradient Sign Method （FGSM），Projected Gradient Descent （PGD），and Carlini-Wagner （C&W） attack， and constructs multi-type adversarial samples， to systematically evaluate the model robustness. Furthermore， by introducing the Mahalanobis distance and utilizing the distribution information in the feature space of the model， a statistical distance based adversarial sample detection method is proposed， enabling effective identification of anomalous inputs. Analysis results of the experiments on a steel surface defect classification task indicate that PGD and C&W attack cause severer damages to AI models， while the proposed method shows good detection capability under different attack conditions， effectively improving the security and robustness of industrial visual classification models.

To address the limited characterization capability of single-source signals in fault diagnosis of slewing bearings， this paper proposes a fault diagnosis method based on acoustic-vibration signal feature fusion and an Improved Plant Rhizome Growth Optimization （IPRGO） algorithm for optimizing XGBoost. In the feature extraction stage， a feature fusion strategy is adopted： for vibration signals， a three-level wavelet packet decomposition is performed to extract energy features from 16 sub-bands as vibration representations； for acoustic signals， after pre-processing such as framing and windowing， 1~12 order Mel-frequency cepstral coefficients （MFCCs） are extracted and combined with four time- and frequency-domain statistical features， including root mean square and kurtosis， to form a 48-dimensional acoustic feature vector. Finally， the 16-dimensional vibration features and 48-dimensional acoustic features are concatenated into a 64-dimensional comprehensive feature vector， constructing a more complete fault state representation system. In the model construction phase， XGBoost is employed as the classifier， and the IPRGO algorithm is introduced to adaptively optimize its hyperparameters. The IPRGO algorithm balances global exploration and local exploitation capabilities by dynamically switching between taproot and fibrous root growth modes， while incorporating a spiral search mechanism to enhance optimization efficiency and avoid local optima. Experimental results demonstrate that the proposed method significantly outperforms models based on single-source signals and traditional parameter optimization methods in terms of diagnostic accuracy and stability for slewing bearing fault diagnosis. This approach effectively overcomes the limitations of single-source signal representation and provides a reliable solution for equipment condition monitoring under complex working conditions.

To address the latency and poor responsiveness of traditional crane inspection methods based on safety devices and periodic manual checks， this study introduces digital twin technology and proposes a structural health monitoring method using an integrated “physical form-virtual behavior-performance mapping” framework. In terms of visualization， multi-disciplinary modeling and system integration enable real-time state perception and physical-virtual synchronization. For rapid mechanical response analysis， proper orthogonal decomposition （POD） and multilayer perceptron （MLP） are combined to reduce high-dimensional physical fields and establish the nonlinear mapping between operating parameters and POD modal coefficients， with full-field reconstruction via inverse POD transformation. Experiments on a 1-ton double-girder gantry crane show that the method provides accurate predictions with high efficiency， achieving millisecond-level high-dimensional physical field prediction and improved real-time monitoring.

Aiming at the fault diagnosis problem of high-speed train wheelset bearings， a combined sparse supervised principal component analysis method is proposed. This method first constructs the combined energy modal space of bearing fault signals using high-order energy operators and multivariate filtering curves. Then， a supervised component model is established with the help of significant segments of the resampled periodic spectrum. Furthermore， sparse supervised principal component analysis is conducted to obtain the output damage principal component under the guidance of the supervised component model. By comparing the theoretical fault characteristic frequency with the prominent frequency components in the square envelope spectrum of the output signal， the method determines whether a fault has occurred in the high-speed train wheel set bearings. The results of experimental data analysis show that the proposed method can diagnose the damage characteristic components from the signals of wheel set bearings in high-speed trains， efficiently determine the damage condition of wheelset bearings in high-speed trains， and provide reference value for practical engineering applications.

Regarding the efficient deployment of steel surface defect detection algorithms， this paper proposes a channel pruning and compression method for the YOLOv8 model. Firstly， L1 regularization is applied to the scaling factors of Batch Normalization layers during training to identify redundant channels. Then， based on a global pruning ratio and layer-wise protection thresholds， channels with scaling factors close to zero are pruned. Furthermore， the pruned compact network is fine-tuned to recover the accuracy loss due to pruning. Comparative analyses are conducted under different pruning ratios. Experimental results on the NEU-DET and NRSD-MN public datasets show that when the pruning ratio is 0.4， the detection accuracy decreases by only 1.41%， while the model parameters are reduced to 51.1% of the original， and GPU inference speed increases by 1.59 times. This compression method significantly reduces memory consumption and computational latency while maintaining high model accuracy， laying a foundation for subsequent deployment on edge devices.

The existing deep learning models have low recognition accuracy and robustness in PCB defect detection， mainly due to three aspects： loss of cross-level feature information， redundant dense attention computation， and insensitivity of multi-scale pooling to small target features. Based on YOLOv8， a PCB defect detection model named YOLO-GDS is proposed. The core improvements and key issues to be addressed are as follows： Firstly， to solve the problem of cross-level information loss， Gold YOLO is introduced to replace the neck network in the original model， and a "collect distribute" mechanism is used to achieve direct information flow between any layers， fundamentally avoiding the problem of information loss in the original structure； Secondly， to solve the problem of high computational cost in dense attention， the deformable self attention module DAttention is used to replace the C2f module in the YOLOv8 backbone network， simplifying the original global dense interaction into sparse computation based on key sampling points， capturing more informative features， improving computational efficiency and accuracy of feature representation， while significantly reducing the computational redundancy of the attention module； Finally， to solve the problem of small-object features being easily obscured in the multi-scale pooling process， the SPPF-LSK module combined with an attention mechanism is used instead of the SPPF module for multi-scale feature extraction. The receptive field is dynamically adjusted through a large kernel selection mechanism， and the attention distribution is adaptively adjusted based on the input data to enhance the model's detection capability for small objects. Different from the method of only replacing a single module， YOLO-GDS realizes more accurate detection of complex defects in PCB through collaborative optimization of three dimensions of "feature-fusion， feature-selection， and feature-enhencement". The experimental results show that， compared with the original model， YOLO-GDS has improved accuracy， recall， mAP@0.5， and mAP@0.5∶0.95 by 0.2， 0.7， 0.3， and 3.2 percentage points respectively. The research results indicate that YOLO-GDS has certain potential in the practical application of PCB defect detection， and can provide strong technical support for improving the efficiency of PCB defect detection.

Aiming at the low accuracy of traditional detection methods for spots， cracks and peeling defects of hand-rolled cigars， an improved PBD-YOLOv5s detection algorithm is proposed. Based on YOLOv5s， the algorithm introduces the BiFPN to enhance multi-scale feature fusion， adds a P2 small-object detection layer to improve the capture capability of tiny defects， adopts a decoupled head to separate classification and regression tasks for reducing task interference， and selects the EIoU loss function to optimize bounding box regression accuracy. Experimental results show that the mAP0.5 of the PBD-YOLOv5s model reaches to 95.3%， which is 16.7% higher than that of the original YOLOv5s model. The most significant improvement in model detection accuracy is achieved through the integration of BiFPN， a P2 small-object detection layer， and a decoupled head， resulting in an mAP0.5 of 95%， which is increased by 16.4% compared with YOLOv5s. The algorithm can effectively improve the detection accuracy of cigar appearance defect detection and provide technical support for improving the appearance quality of cigars.

The "black-box" nature of deep learning-based point cloud segmentation models limits their reliable application in safety-critical fields such as power line inspection. To address the challenges of low explanation resolution and detail loss in existing point cloud interpretability methods， this paper proposes a generic interpretation framework based on hierarchical feature fusion. The framework generates initial heatmaps by extracting and fusing gradient weights and feature maps from multiple network layers. It then employs a precise upsampling method based on KD-tree indexing to map these heatmaps back to the original point cloud. Finally， an adaptive weighting strategy is designed to integrate information from different layers， producing a high-resolution heatmap. Experimental results on transmission tower point cloud data demonstrate that our method achieves 92.3% accuracy in component localization， representing a significant improvement over traditional methods， particularly in identifying fine components such as insulator strings and grading rings. Moreover， it maintains stable performance under complex conditions， including 500 kV double-circuit lines on the same tower and vegetation occlusion scenarios， confirming its effectiveness and robustness in practical power inspection applications. This study provides a novel technical pathway for interpretable point cloud analysis in the power industry and offers essential support for the development of intelligent inspection systems.

To address the issues of low global obstacle avoidance efficiency and poor environmental adaptability in traditional intelligent algorithms applied to mobile robots in complex scenarios， a Biodiversity Feedback Strategy-based Improved Crayfish Optimization Algorithm （BF-ICOA） is proposed for mobile robot path planning and obstacle avoidance. The ICOA algorithm is employed for global path planning， utilizing multi-crayfish population metrics to stimulate the generation of population diversity. During the foraging phase， a novel foraging update strategy enhances the population's competitive optimization capability. A new biodiversity feedback strategy is introduced to trigger the exploration depth of global paths and stimulate inter-individual competition under varying environmental conditions. A path refinement strategy is proposed， utilizing an enhanced error metric to identify and eliminate redundant paths. Experimental validation is conducted using diverse fitness functions and multi-scale grid maps. Results demonstrate a 20.17% reduction in path planning length and a 2.2-second decrease in inference time compared with the original algorithm， validating the reliability of the improved method for mobile robot trajectory planning.

Aiming at the energy efficiency optimization of 7-DOF redundant wiring robots in the automated wiring of aviation wire harnesses， a 7-axis coordinated motion planning method is proposed to solve the problems of high energy consumption of large-inertia guide rails and susceptibility to motion singularities during movement. Firstly， the RRT-Connect algorithm is applied to generate a discrete global path， cubic B-spline curves are adopted for path smoothing， and inverse kinematics solution is completed based on the principle of differential kinematics. Secondly， a weighted generalized inverse non-uniform velocity distribution framework is established， which prioritizes motion demands to the six rotational joints and reduces the motion participation of the seventh axis to achieve system energy efficiency optimization. In addition， an artificial potential field function with saturation constraints is introduced in the null-space to guide the mechanism back to the neutral position at the midpoint of the seventh axis stroke， avoiding limit singularities. Meanwhile， a closed-loop feedback correction mechanism based on Lyapunov stability is introduced to ensure trajectory tracking accuracy. Finally， simulation results verify that the proposed motion planning method yields smooth trajectory curves， reduces the stroke of the seventh axis， improves system manipulability， lowers the utilization rate of the robot’s large-inertia axis， and enhances the configuration flexibility of the system while ensuring end-effector accuracy.

In recent years， research on lower-limb exoskeleton robots has garnered increasing attention， with load-bearing and mobility-focused exoskeletons emerging as a key research direction. However， current lower-limb load-bearing exoskeletons still suffer from issues such as poor support and assistance efficiency， rigid human-exoskeleton interaction， and significant interference with the human body， limiting their performance and practical application. This paper proposes an iterative gait generation strategy and adaptive impedance assistance method based on human-exoskeleton end point alignment. This approach automatically adapts to individuals with varying body types and generates matching trajectories according to distinct human gait characteristics. Experimental results demonstrate that the exoskeleton can support walking with loads exceeding 30 kg while offloading at least 62.71% of the body’s pressure， and reduce human metabolism by no less than 50.62%， effectively validating the correctness and practicality of the control system. The proposed recognition and control method requires only lower-limb pose data and motor encoder information， eliminating reliance on force sensors. Its broad applicability and low development cost enable widespread use across various exoskeleton robot types， demonstrating significant engineering value.

To address the issues of low path planning solution efficiency and difficulty in adapting to dynamic environmental changes due to the complex inspection environment and numerous task points in traction substations， a hierarchical path planning method integrating generative diffusion models and incremental search is proposed. In the upper-level task-point sorting planning， a diffusion model based on matrix encoding networks is constructed to generate a global initial solution for traversing multiple target points through the reverse denoising process， and the 2-opt algorithm is combined for local neighborhood search and optimization. In the lower-level obstacle avoidance path planning， the D^*Lite algorithm is used to generate specific navigation paths， and an incremental cost update mechanism is adopted to achieve online re-planning for sudden obstacles. At the end， Bézier curves are used to smooth the paths to meet the kinematic constraints of the robot. MATLAB simulation results show that in a large-scale scene with 200 task points， the proposed method improves the solution efficiency by more than five times compared with traditional heuristic algorithms， shortens the path length by approximately 10%， and controls the re-planning time within 1.5 ms， enhancing the inspection efficiency and safety of substation robots. The method has also been verified on actual robots.

Aiming at the core demand of long-endurance ocean observation for the continuous service capability of wave gliders， and to solve the fatigue failure problem of the spring assemblies in its adaptive hydrodynamic mechanism under the highly corrosive marine environment and long-period alternating loads， this paper carries out research on the fatigue characteristics analysis and life evaluation of the spring assemblies. Based on the power conversion mechanism of wave gliders， a moment-balanced force model of the spring assembly is established， the load calculation method under wave excitation is derived， and the structural design and material selection are completed. The finite element method （FEM） is adopted to carry out static strength analysis and fatigue life simulation， and the stress distribution， dangerous sections and life distribution law of the spring assemblies are obtained. The test scheme is designed in accordance with national standards， and the accuracy of the simulation results is verified through accelerated fatigue tests. The results show that the maximum working equivalent stress of the spring assembly is 422.13 MPa， which meets the material strength design requirements； the minimum fatigue life predicted by simulation reaches 9.059 million cycles， and the spring assembly prototype has no fracture or plastic deformation after 9 million cycles， with an elastic loss rate of less than 8%， and the error between simulation and test is less than 4%， which satisfies the continuous service requirements of more than 12 months for offshore unmanned platforms. This research can provide a theoretical basis for the anti-fatigue design of elastic components of wave gliders， and also provide an engineering solution for the high-cycle fatigue life evaluation of marine equipment.

To address the problems in the “last-meter” material handover between AGVs and process equipment in flexible smart manufacturing lines， including long communication links， local coordination relying on central scheduling， insufficient stability of wireless communication in strong electromagnetic environments， and non-unified interface specifications， a direct optical communication system based on Parallel I/O Optical （PIO） was designed. The system adopts a layered decoupling architecture to separate upper-level task scheduling from real-time end-point handover control. Infrared through-beam sensors are used to establish a bidirectional 8-bit optical communication link between AGVs and process equipment， and hardware-level parallel state encoding is realized through discrete I/O mapping of PLCs. Meanwhile， mechanical， electrical， and optical interface specifications are standardized， and four standardized operation modes are defined. Practical application results show that the one-way PIO communication time is stable within 20 ms. Under a strong electromagnetic interference environment， the system operated continuously for 72 h without status misjudgment or communication interruption， and more than 2000 docking operations were completed without communication failure， with a PIO optical communication success rate of 100%. The proposed system improves the real-time performance， determinism， and reliability of end-point material handover， and provides a reusable reference for direct coordination of heterogeneous equipment in flexible production lines.

Aiming at the industrial bottlenecks in traditional conveyor idler production lines， including high labor dependence， isolated information silos among different processes and inadequate closed-loop correction capabilities， and to meet the stringent requirements for high-precision and high-quality in conveyor idler manufacturing， this paper studies an intelligent conveyor idler production line based on error closed-loop control technology. Core to the proposed line is the construction of a collaborative architecture comprising the application service layer， data layer， algorithm model layer and equipment execution layer by virtue of an intelligent management system. Driven by real-time detection and feedback information， the full links of precise measurement， error calculation， decision issuance and closed-loop control are seamlessly integreted. Meanwhile， supported by the precise tolerance control of conveyor idler pitch and the detection and control mechanism for multiple performance parameters， an intelligent conveyor idler production line with functions of error closed-loop control and self-optimization is developed. Verified by engineering practices， the product qualification rate of conveyor idlers manufactured by this intelligent production line reaches 100%， and the superior product rate hits 94.6%. This study achieves a leap of conveyor idler manufacturing from semi-automatic and automatic modes to the intelligent manufacturing mode with error closed-loop control， and provides technical support and engineering cases for the intelligent upgrading of production lines for coal mining machinery and equipment manufacturing.