ISSN 1674-5949 CN 31-2023/U
A large amount of process knowledge accumulated during industrial programmable logic controller （PLC） development exists in tacit form， embedded in historical control logic and dependent on engineers' experiential judgment， making it difficult for large language models to directly utilize. To address this problem， this paper proposes a systematic engineering method that transforms tacit process knowledge into reusable process templates with hierarchical structure and control topology details， through a dual-layer knowledge graph schema and a bidirectional fusion construction strategy. Using PLC code generation as a downstream evaluation task， we target two structural barriers — knowledge accessibility and functional association completeness — and validate the effectiveness of the proposed method through a knowledge granularity gradient experiment and a graph-versus-vector retrieval recall comparison experiment， respectively. Experimental results demonstrate that， on code-segment-level tasks， the process templates automatically extracted by the proposed method approach the performance of manually written pseudocode-level fine-grained knowledge specifications in terms of control requirement information supply； on functional-block-level tasks， the graph-structured knowledge organization successfully mitigates the systematic deficiency of vector retrieval in preserving functional association completeness.
Industrial data faces challenges such as complex entity structures， severe nesting， and high levels of noise interference， leading to inaccurate boundary detection and lack of hierarchical dependencies in existing methods. This paper proposes an improved Multi-Granular Adversarial Named Entity Recognition model， MG-AdvNER. First， to address the issue where traditional adversarial training blurs entity boundaries， we propose an Adaptive Boundary-Aware Adversarial Training （ABAT） strategy. ABAT dynamically adjusts perturbation weights at boundary positions， enhancing model robustness while maintaining boundary sensitivity. Second， we construct a dual-layer CRF structure and design an Inter-layer Constraint Propagation mechanism to explicitly model the logical dependencies between coarse-grained and fine-grained labels， resolving inconsistency issues in multi-granular predictions. Furthermore， we constructed an industrial dataset containing 17，000 entities and conducted extensive experiments on both this dataset and public datasets. The results demonstrate that MG-AdvNER achieves an F1 score of 92.20% on the industrial dataset， significantly outperforming baseline models， which validates the effectiveness and generalization capability of the method in complex industrial scenarios.
To address insufficient feature extraction from bearing vibration signals， a bearing fault diagnosis method based on multi-scale time-frequency joint channel attention mechanism and multi-head self-attention mechanism （MSTFCA-MHA） is proposed， effectively enhancing multi-scale feature extraction capability. Firstly， wide-kernel convolution is employed to process normalized signals and expand the receptive field of the network. Secondly， a time-frequency joint channel attention module is constructed to extract and fuse multi-scale local features in both time and frequency domains through multi-scale parallel convolutions. Thirdly， multi-head self-attention is utilized to model global dependencies， further strengthening feature extraction capabilities. Finally， experiments on the Paderborn University bearing dataset and stacker bearing dataset demonstrate that， compared with WDCNN-LSTM， MACNN and DRSN-Transformer， the proposed method achieves higher average recognition accuracy of 99.84% and 100% respectively， while its noise resistance performance is validated under various signal-to-noise ratio conditions.
With the rapid advancement of the manufacturing industry， anomaly detection in complex assembly scenarios has emerged as a critical component for ensuring product quality and production safety. Traditional detection methods， often constrained by limitations in detection accuracy， efficiency， and adaptability， struggle to meet the stringent requirements of modern manufacturing. To address these challenges， this paper proposes an anomaly detection method for complex assembly scenarios based on simulation-to-reality comparison. This approach aims to achieve accurate identification and efficient management of anomalies during the assembly process by integrating virtual simulation data with real-world detection data. Experimental results demonstrate that in typical anomaly scenarios—such as missing components， misassembly， and foreign object intrusion—the proposed method achieves an average improvement of 22.1% in the F1 score and over 20% in the AUC metric compared with the state-of-the-art unsupervised methods， exhibiting superior detection accuracy and robustness. Furthermore， comparative experiments validate the distinct advantages of this method over existing zero-shot anomaly detection techniques.
To address issues such as the high proportion of small objects， insufficient feature representation， and interference from complex reflective backgrounds in the detection of bubble defects in wing bonding areas， we propose an improved YOLO model， HAM-YOLO （Hybrid Attention Mechanism YOLO）. CBAM and SE attention modules are introduced into the backbone network to enhance channel and spatial feature representation capabilities； dual attention and SE attention mechanisms are embedded in the C2f module of the neck network to improve the fusion and selection capabilities of multi-scale features； simultaneously， a 1×1 convolutional layer is added during the feature fusion stage to optimise channel information interaction. Furthermore， an improved loss function combining GHM and dynamic IoU is designed to enhance the model’s capability to focus on difficult samples and improve boundary regression accuracy. Experimental results demonstrate that the proposed method outperforms baseline models and various alternative approaches in terms of detection accuracy and comprehensive performance metrics on a self-built aircraft wing bonding defect dataset， validating its effectiveness and superiority in complex industrial scenarios.
To address the decline in detection accuracy and insufficient real-time performance caused by densely stacked profiles， large illumination variations， and limited computational resources of embedded devices in industrial manufacturing and warehouse logistics scenarios， an improved YOLOv11n-based embedded profile-counting detection model is proposed in this paper. To adapt to diverse poses， complex lighting conditions， and varying backgrounds on industrial sites， we construct an industrial profile dataset composed of real-scene captures and Blender-rendered samples to improve the model’s cross-scene generalization. The YOLOv11n is optimized along multiple dimensions： a DBB multi-branch convolution block is introduced into the backbone to enhance multi-scale detail representation； BiFPN replaces Concat in the neck to achieve efficient cross-scale feature aggregation； the detection head integrates a CBAM attention mechanism and a CDetect decoupled dual-branch structure to improve localization accuracy for overlapping objects； WIoUv3 is adopted in the loss design to optimize bounding-box regression in dense scenarios. Experimental results show that the improved model achieves 98.75% mAP on the test set with only 2.98 M parameters. Compared with mainstream models such as Faster R-CNN and YOLOv8n， it performs better on detecting densely overlapping profiles and fine-grained structures while substantially reducing inference latency. The proposed method balances high accuracy and real-time performance， meeting the requirements of embedded industrial profile automatic counting and providing an efficient and reliable solution for intelligent manufacturing vision systems.
As power conversion devices， solid-state pulse modulators， leveraging their advantage of compact design and layout， are generally integrated with pulse transformers to serve as the mainstream technical solution for the discharge system of miniaturized electronic linear accelerators. In practical operation and maintenance， traditional fault detection and localization solutions require the system to be forced to shut down after a fault occurs in the solid-state modulator， with on-site diagnosis conducted by maintenance personnel—this results in the interruption of production processes. Meanwhile， manual fault troubleshooting is limited by strong reliance on experience and high labor costs； furthermore， the compact layout of the equipment further increases the difficulty of implementing fault localization. Aiming to solve the bottlenecks of traditional manual maintenance， this paper proposes an online real-time faulty unit localization method based on adjusting the pulse formation timing of energy storage units. In addition， to adapt to different user requirements and load conditions， single-pulse and multi-pulse fault localization methods are developed， which provide accurate and efficient technical support for maintaining accelerator discharge systems.
To address the delay-induced bandwidth bottleneck in the current loop of three phase permanent magnet synchronous motor （PMSM）， this paper investigates how the sampling/update delay， hardware delay， and feedback-filter delay affect the achievable system bandwidth， and establishes a current-loop model that accounts for multi-source delays. Based on the loop modeling， the inverse relationship between bandwidth and the total loop delay is analyzed， and the control-time limitation caused by conventional coordinate transformations and computationally intensive modulation under high control frequencies is clarified. To mitigate the insufficient control time in high-bandwidth scenarios， a phase-current-regulation-based current-loop control architecture is proposed. Combined with an optimized space-vector pulse-width modulation （SVPWM） algorithm， the proposed method omits Clarke and Park transformations and avoids trigonometric and division operations， thereby significantly shortening the control path and improving computational efficiency. An experimental PMSM drive platform is built， and frequency-domain tests with white-noise excitation are conducted to identify and validate the dynamic characteristics of the proposed strategy. Experimental results show that， at a control frequency of 400 kHz， the implemented current loop achieves a single-cycle execution time of approximately 1.07 μs and an open-loop 0 dB crossover frequency of about 11.2 kHz； near this frequency， the closed-loop phase attenuation is less than 70°. These results indicate that the proposed method can realize high-bandwidth current-loop control under stringent timing constraints at high control frequencies.
The super twisting sliding mode observer （STSMO） used for position-sensorless control of permanent magnet synchronous motors （PMSM） is only applicable to medium and high speeds. Moreover， its convergence speed is slow in the face of strong disturbances. Therefore， this paper proposes a position-sensorless control scheme of PMSM based on combination of IF control and adaptive linear super twisting sliding mode observer （LSTSMO）. In low speed region， IF control with open-loop rotational speed and closed-loop current is used. The phase relationship between virtual synchronous frame and rotor synchronous frame is analyzed. In high speed region， the STSMO is modified for position-sensorless control. A linear term is introduced to accelerate the convergence rate. By designing a reasonable adaptive rate of parameters， the observer parameters change with the variation of the observed rotational speed， thus extending the operating speed range. PMSM has poor anti-interference capability and is prone to losing synchronization. Therefore， this paper designs an extended state observer （ESO） to measure the load torque and compensate it into the q-axis current. The smooth transition of the two control schemes is achieved by rationally designing a proper transition process based on the phase relationship between the double dq space. Finally， the effectiveness of the proposed composite strategy is verified through simulation and experiments.
The increasing penetration of renewable energy and the expanding scale of DC transmission have become major development trends in modern power systems. To determine the appropriate integration ratios for renewable energy and DC transmission， it is essential to first clarify the impact mechanism of their combined effects on the transient voltage stability of the system and to investigate the dynamic evolution trends of the system's stability boundaries， thereby providing decision support for the progressive planning and risk early warning of the power grid. This study focuses on key trends in power system evolution， establishing a system model and developing transient voltage stability indicators applicable to multiple scenarios. It further quantitatively analyzes the variation patterns of transient voltage stability under the synergistic evolution of renewable energy and DC penetration ratios. By improving and applying the polynomial approximation method based on collocation points to accurately calculate critical stability boundaries， the dynamic evolution trends of the system's transient voltage critical stability boundaries are revealed. The results indicate that the synergistic increase in renewable energy and DC penetration ratios significantly weakens the system's transient voltage stability， with a notable downward shift in the critical stability boundary.
With the development of green manufacturing and precision manufacturing， in the face of the complex curved surface characteristics of aviation， near-net-shape forming processes such as precision casting and precision forging are widely used to obtain small allowance blanks. However， due to the lack of materials in some local areas， the processing may be substandard. To address this issue， a multi-level allowance optimization method under contour tolerance constraints is proposed. It changes the concept of rigid displacement of the workpiece and allows the design section lines to change their relative positions or deforms within the contour tolerance zone respectively， thereby increasing the feasible domain space for optimization and enhancing the capability to find the target machined surface. Furthermore， after the optimization of the allowance， the tolerance zone space was occupied， and a method for adaptively adjusting the cutting path was proposed. Finally， the effectiveness of the proposed allowance optimization and cutting path adjustment methods was verified through the machining of typical complex surface featured blade bodies.
In automated TIG welding systems for medium-thick plate， accurate and stable perception of weld seam geometry is critical for achieving high-quality welds. To address the limitations of conventional methods， such as poor feature adaptability， weak generalization capability， and low robustness in complex working conditions， this paper proposes a two-stage feature extraction algorithm integrating projection range and energy constraints. The method begins by constructing a projection range metric combined with adaptive threshold segmentation to achieve robust initial extraction of feature points. Subsequently， an optimization model incorporating normal attraction and repulsive potentials is developed to refine the positions of the feature points and improve distribution homogenization. Finally， a local normal vector statistics-based approach is employed to accurately identify weld valley points. Experimental results demonstrate that the proposed algorithm achieves an average extraction accuracy of less than 0.35 mm and F1-scores above 0.83 on typical medium-thick plate weld seams. It significantly enhances adaptability to different weld seam types and anti-interference capability， providing a stable and reliable feature perception solution for intelligent welding systems.
The impact of press-fit damage on the high-temperature service performance of assembly stuck points was often overlooked by traditional empirical tolerance allocation methods. This led to untargeted maintenance and high resource consumption in the production environment of solar silicon film PECVD deposition. Finite element simulation models of the stuck point press-fitting and high-temperature service process were established. The influence of varying interference levels and environmental temperatures on press-fit performance damage was analyzed. Numerical simulations show that contact stress at the stuck points increases significantly during the initial press-fit stage， while residual stress escalates with increasing interference magnitude and ambient temperature. Taking a 15 μm interference fit as an example， when the temperature rises from 400 ℃ to 600 ℃， the residual stress increases by approximately 45%. It is recommended to appropriately reduce the interference under high-temperature conditions to enhance assembly stability. A B-spline interpolation method was employed to establish the mapping relationship among boat-piece hole diameter expansion， interference level， and initial aperture. Based on this analysis， a condition-based maintenance method for stuck points under different temperature conditions was developed using a Weibull distribution. A preliminary case study indicates that， compared with the empirical 5 μm compensation method， the proposed condition-based maintenance approach limits the variation of stuck point contact stress to less than 10% and can reduce the maximum cumulative damage parameter of the boat piece from 15.2 μm to 11.9 μm. This provides a technical reference for silicon film deposition production and the scientific maintenance of graphite boat performance.
The accuracy of vehicle dynamic model parameters plays a critical role in high-performance vehicle motion control systems， particularly under high-speed driving and near-limit adhesion conditions， where even minor parameter deviations can significantly degrade control performance. Conventional offline parameter identification methods based on steady-state experiments rely heavily on specific test environments and operating conditions， making them difficult to apply in real-world scenarios that require real-time adaptability. To address this issue， this paper investigates online parameter identification for vehicle motion control and proposes a hybrid online identification framework that integrates physics-based models with data-driven techniques. The proposed method relies solely on data collected during normal vehicle operation and employs a neural network to learn and compensate for the residual errors of the physics-based model， thereby significantly enhancing identification robustness while preserving physical consistency. Tire parameters of the Pacejka model are selected as the identification targets， and the proposed approach is systematically validated through simulation studies. Experimental results demonstrate that， under noise interferences and model mismatch conditions， the proposed method achieves higher identification accuracy and stability than conventional online nonlinear least-squares approaches， while also delivering notable improvements in control performance during real-world vehicle testing.
To address the issue that the excessive self-weight of cross-span power transmission line inspection robots compromises their stability during obstacle-crossing， a lightweight design of the robot's walking arm is investigated. Firstly， the force characteristics of the walking arm under dual-arm and single-arm line-grasping postures are analyzed to establish the mapping relationship between the arm's mass and the body offset during obstacle negotiation. Secondly， a virtual prototype model of the inspection robot is developed in kinematic simulation software to simulate the obstacle-crossing process， thereby obtaining the load conditions of the walking arm under extreme operating scenarios. Subsequently， variable density topology optimization is performed on the moving support and shell of the walking arm and telescopic arm. Comparative finite element analysis （FEA） of the maximum stress and deformation before and after optimization indicates that the static strength and stiffness of the optimized arm satisfy the design requirements. Notably， the mass is reduced by 1.38 kg， achieving a weight reduction ratio of 28.59%. Finally， kinematic simulations and prototype experiments are conducted to verify the body deflection angle of the light-weighted inspection robot during cable-guided obstacle-crossing process. The results demonstrate that the lightweight design reduces the maximum body offset angle， leading to superior obstacle-crossing stability.
This paper investigates a vision-based dual-arm peg-in-hole assembly method for a humanoid dual-arm robot platform equipped with dexterous hands. The research addresses the requirements for grasping adaptability and motion anthropomorphism in natural environments. This paper uses steel pipes and pipe caps as assembly objects， employing the YOLOv8n-seg model to detect and generate segmented images； Combines depth images with the FoundationPose model to determine the pose of parts in 3D space， enabling visual recognition and positioning of assembly components. We further propose a pose alignment method that extracts part axes based on recognition results， calculates the target pose of the end-effectors of the robotic arms， and generates corresponding motion waypoints to accomplish vision-based grasping and peg-hole alignment. During the assembly phase， we introduced a helical hole-finding strategy to achieve compliant assembly of peg and hole components to complete the entire dual-arm assembly process. The entire system is integrated using Robot Operating System， achieving full automation from visual perception and grasping alignment to peg-hole insertion.Experimental results demonstrate that the system can effectively and consistently complete the dual-arm peg-in-hole assembly task.
With the growing diversity and personalization of user demands， as well as continuous technological advancements in manufacturing， mass customization has become a prevailing trend. In mass customization， product configuration plays a critical role in enabling rapid design of personalized products. Existing product configuration methods generally consider user requirements， enterprise profitability， and low-carbon design， but rarely focus on evaluating product circularity performance. In the context of green， low-carbon， and circular development， a product configuration model is proposed that incorporates "product circularity index" into the optimization objectives. This model calculates the circularity index of product modules and establishes a multi-objective optimization model with customer satisfaction， cost， and product circularity as the optimization objectives. The Pareto-optimal solution set is obtained using a Non-dominated Sorting Genetic Algorithm. A case study on an electric vehicle produced by a well-known automotive company is conducted to verify the feasibility and effectiveness of the proposed module configuration scheme.
In order to solve the problem of production progress prediction in the equipment manufacturing process， a production progress prediction method based on Spark and model tree is proposed by utilizing the big data resources of the manufacturing system. The method provides a logical process and system framework for predicting production progress， and analyzes the elements and influencing factors of production progress. Then， the influencing factor features of production progress are selected by using the ReliefF feature selection algorithm to remove redundant and irrelevant features. Furthermore， the model tree algorithm is distributed based on the Spark framework， and used to construct and train a production progress prediction model to achieve accurate and real-time prediction of future progress. Finally， the effectiveness and feasibility of the proposed method are verified by an example of satellite processing.
To optimize the processing quality and efficiency for honeycomb materials， the ultrasonic cutting tool holder for a serrated disc blade was designed. The acoustic system of the ultrasonic cutting tool for the serrated disc blade was designed using the four-terminal network method， and the relationship among the number of teeth， resonant frequency， and amplitude was derived. Modal analysis was also conducted. The vibration mode diagrams indicate that as the number of teeth increases， the resonant frequency decreases， and the amplitude increases accordingly. An impedance analysis test on the acoustic system of the ultrasonic cutting tool for the serrated disc blade revealed a theoretical-to-actual error of 3.9%， which meets machining requirements. Ultrasonic cutting experiments on honeycomb materials using the ultrasonic tool holder were conducted， demonstrating the feasibility of applying the ultrasonic cutting tool holder for the serrated disc blade in production.
