To address the challenges of inner-outer loop compliant control， severe force coupling， and grasping instability caused by parameter uncertainties in dual-arm closed-chain cooperation， this study proposes a master-slave-architecture-based inner-outer-loop dual-channel compliant control framework. First， key object parameters such as mass and center of mass are precisely identified online via a multi-pose statics method， which effectively improves the accuracy of the dynamics model and lays the foundation for accurate force decoupling and compensation. Furthermore， considering the differences in contact modes during practical operations， an analysis of grasping tightness constraints is introduced. By constructing a pseudo-inverse load distribution model with a modified weight matrix， physical boundary conditions such as unidirectional forces are incorporated into the solution process. Utilizing the grasp matrix’s null-space projection， the generalized end-effector forces are then decoupled into "load-bearing forces" （for balancing external loads） and "pure internal forces" （for maintaining grasp stability）. In this architecture， the slave arm dominates the inner force-control loop to actively suppress gripping force fluctuations， while the master arm governs the outer admittance-control loop to compliantly respond to environmental interactions. Experimental results demonstrate that the method effectively decouples the inner- and outer-loop control channels， significantly enhancing the stability， safety， and environmental adaptability of dual-arm collaborative manipulation， thus providing a reliable technical solution for high-precision closed-chain operations under complex contact conditions.

To address the limitations of existing rolling bearing damage diagnosis methods， including reliance on manual feature extraction， lack of semantic interpretability， and insufficient generalization under complex operating conditions， this paper proposes a method for rolling bearing damage identification and interpretable diagnosis based on adaptive frequency-band feature modeling and a hybrid CNN-Transformer network， incorporating a large language model as an auxiliary semantic interpretation module. First， the raw vibration signal is transformed into time-frequency representations using short-time Fourier transform. Combined with the distribution of bearing characteristic frequencies， an adaptive frequency-division feature modeling method is proposed to divide the signal into multiple physically meaningful frequency-band subspaces， enabling fine-grained representation of different damage-related frequency components. Second， a feature extraction network integrating convolutional neural networks and Transformers is designed to perform deep mechanical feature learning and cross-band information fusion on the segmented time-frequency signals， thereby enhancing the weak damage feature identification capability of the model. The fused deep features and key frequency-band responses are then converted into structured semantic representations. During the diagnostic result generation stage， the model’s output—including damage categories， key frequency-band responses， and feature contribution information—is organized into a structured semantic representation， and fed into a large language model for auxiliary interpretation. Meanwhile， by incorporating typical bearing damage characteristic frequencies， a mapping between frequency bands and damage mechanisms is established， improving the interpretability and engineering credibility of the diagnostic results. Finally， experiments conducted on public bearing datasets demonstrate the effectiveness of the proposed method. Compared with various deep learning models， the proposed approach achieves superior performance in both identification accuracy and interpretability.

Significant progress has been made in natural language to SQL （NL2SQL） technology， driven by large language models （LLMs）. However， when deployed in specific vertical domains， accuracy often drops sharply due to implicit business rules and term ambiguities. Although existing retrieval-augmented generation （RAG） methods can introduce external knowledge to achieve some improvements， they still suffer from several issues： lack of importance discrimination among knowledge snippets （i.e.， no confidence scores）， waste of context due to redundant knowledge （semantic redundancy）， misleading LLMs with contradictory knowledge （logical conflicts）， and inability to evolve through interaction （static knowledge base）. To address these challenges， we proposed a knowledge-driven adaptive method for NL2SQL （KD4SQL）. First， the method employs an LLM-aware dynamic knowledge re-ranking strategy that integrates retrieval scores with usefulness scores to achieve importance-based sorting. Second， a confidence-driven semantic redundancy elimination mechanism automatically removes duplicate snippets. Third， by incorporating an applicability-condition-aware contradiction resolution mechanism， it parses conditions and dynamically handles conflicts. Finally， a feedback-driven knowledge evolution strategy incrementally updates weights and extracts new rules based on user corrections or execution failures， thereby dynamically expanding the knowledge base. Validated on real-world manufacturing data， KD4SQL achieves an execution accuracy of 82.7%， outperforming all baseline models. In terms of time overhead， the average end-to-end latency for a regular query is 5.2 seconds， which represents an increase of approximately 27% over the best-performing baseline； however， the accuracy gains far outweigh this additional inference cost. Notably， since the proposed innovations are decoupled from the model training process， the method can be rapidly transferred to other domains with strong adaptability.

To address the inefficiency of operation and maintenance （O&M） caused by complex working conditions， dispersed knowledge， and experience-dependent fault diagnosis for gantry cranes in shipyard dock areas， this paper proposes a trinity of intelligent O&M approach that integrates Knowledge Graph （KG）， Large Language Model （LLM）， and operational-condition-data-based equipment prognostics and health management （PHM）. A five-dimensional O&M knowledge graph encompassing structure， faults， maintenance， records， and status is constructed to enable structured representation and dynamic updating of multi-source heterogeneous knowledge. A collaborative reasoning mechanism based on Knowledge-Graph-enhanced Retrieval-Augmented Generation （KG-RAG） is designed， which injects the logical constraints from the knowledge graph and the dynamic indicators evaluated by PHM—such as actual load spectrum coefficients and remaining working cycles—into the LLM， effectively suppressing hallucinations and generating traceable diagnostic reports. A remaining useful life assessment method based on actual operational condition data is introduced to achieve dynamic quantification of overall equipment health and real-time updating of the knowledge graph. Practical case validation demonstrates that the proposed method significantly improves the accuracy of fault root cause localization and the interpretability of maintenance decisions， while reducing the risk of unscheduled downtime， providing a reproducible technical paradigm for "knowledge-data" dual-driven intelligent O&M of complex equipment.

Currently reported weld defect detection methods suffer from issues such as missed detection of small-scale object defects and insufficient multi-scale feature representation.So， this paper proposes a YOLOv-RPW-based weld defect detection method. First， the method introduces a C3K2-RAFConv module based on receptive field attention mechanism into the backbone network of the YOLOv11 model， enhancing local receptive field feature representation to improve the extraction of fine-grained defect features like cracks and pores. Second， considering the high proportion of small-scale objects in weld defects， a lightweight P2-Lite small-object detection branch is designed in the detection head to fully utilize shallow high-resolution features， boosting small defect detection capability while controlling model complexity. Finally， the WIoU loss function is introduced to optimize the bounding box regression process， enhancing the model's localization accuracy for defect objects. Experimental results demonstrate that the proposed YOLOv-RPW model achieves superior performance in weld defect detection tasks， with P， R， mAP@0.5， and mAP@0.5：0.95 reaching 91.26%， 83.58%， 90.31%， and 67.24%， respectively， showing improvements of 4.31%， 4.31%， 4.35%， and 3.27% compared with the baseline YOLOv11 model. In summary， the proposed method effectively reduces false and missed detections of weld defects， significantly improves the model's detection accuracy and localization capability for small-scale weld defects， and provides an effective solution for automated weld defect inspection in industrial scenarios.

To address missed and false detections caused by small objects （such as pedestrians and vehicles）， complex backgrounds， and dense occlusion in UAV-based inspection of intelligent manufacturing parks， this paper proposes an aerial small object detection algorithm named EGA-YOLO. Based on YOLOv8， EGA-YOLO introduces the enhanced channel-contextual attention （ECCA） module to enhance feature representation， an enhanced multi-scale fusion module "EGD-Fusion" to reduce feature fusion loss， and the C2fGhost module to reduce redundant features. Experiments are conducted on VisDrone2019-DET dataset. The results show that EGA-YOLO achieves mAP@0.5 and mAP@0.5：0.95 of 44.9% and 30.1%， respectively—an improvement of 8.1 and 8.3 percentage points over YOLOv8s—validating its effectiveness and generalization capability.

Aiming at the problems of low defect edge contrast， wide defect scale span and dense tiny defects in online steel surface quality inspection， which cause traditional lightweight models to suffer from insufficient recognition accuracy and poor adaptability to real-time production line working conditions， this paper proposes an improved steel surface defect detection algorithm named APGS-YOLO based on YOLO11. Firstly， a C3k2-AP multi-scale edge enhancement module is designed to improve the feature extraction capability for tiny defects and suppress the interference from steel plate background textures by means of feature branching and multi-directional asymmetric convolution. Secondly， the GhostBottleneckV2 lightweight attention residual module is introduced to enhance global feature fusion while reducing model computational overhead. Thirdly， a cross-scale feature fusion strategy is adopted to strengthen the extraction capability of multi-scale defect features. Finally， a CSSDIoU scale-sensitive hybrid loss function is constructed to improve the localization accuracy of tiny defects. Experimental results show that the mAP@0.5 and mAP@0.5：0.95 of the proposed APGS-YOLO reach 80.3% and 48.2%， which are 2.9% and 2.5% higher than those of the original YOLO11n. The proposed model outperforms other mainstream object detection models. Meanwhile， the model achieves a 7.0% parameter compression rate with an inference speed of 86 frames per second. The proposed algorithm balances detection accuracy and real-time inference performance， which can satisfy the automatic online quality inspection requirements of steel production lines.

To address the issues of strong subjectivity in feature extraction and insufficient accuracy and generalization under extreme operating conditions in traditional fault line selection methods for distribution networks， this paper proposes a novel fault line selection approach that combines electrical iris-spectrograms with a dual-channel convolutional neural network （DCCNN）. First， an adaptive denoising and smoothing model is employed to preprocess zero-sequence current signals， thereby enhancing the signal-to-noise ratio of features. Second， the denoised signals are mapped into iris-spectrograms using the Short-Time Fourier Transform （STFT）. Through a circular plot topological structure—where azimuth represents time and radius represents frequency—the signal's periodic physical characteristics is explicitly integrated. Furthermore， a superpixel segmentation algorithm is introduced to perform semantic clustering on homogeneous energy regions in the time-frequency domain. This aims to remove redundant interference information by restricting the visualized dynamic range， while simultaneously strengthening the morphological characteristics of the fault energy envelope. Finally， a parallel DCCNN is constructed to simultaneously mine the one-dimensional raw sequence features of the zero-sequence current and the two-dimensional image topological features of the iris-spectrogram， achieving precise fault line identification through multi-dimensional feature fusion. Experimental results demonstrate that this method eliminates the dependency of traditional line selection on subjective features and exhibits excellent robustness and noise immunity under extreme conditions， such as high-resistance grounding and communication delays.

Buck-type DC-DC converters are widely used in fields such as aerospace， automotive electronics， and communication power systems， where accurate dynamic modeling under complex operating conditions is of significant importance for system design， performance evaluation， and online analysis. Conventional physics-based models rely on complete circuit parameters and suffer from high computational cost in high-resolution simulations. Pure data-driven methods lack explicit physical constraints and their prediction capability is limited under varying operating conditions. To address these issues， an energy-conservation-constrained physics-informed neural network（EC-PINN） model is proposed to simulate the time-domain response of a closed-loop Buck converter under different operating points， parameter deviations， and load disturbances. To validate the effectiveness of the proposed method， a physics-informed fully connected neural network（FCNN） and a purely data-driven long short-term memory network （LSTM） are established as comparison models. Experimental results demonstrate that EC-PINN achieves higher prediction accuracy for both output voltage and inductor current compared with FCNN and LSTM. Compared with FCNN， the RMSE values of output voltage and inductor current are reduced by 67.9% and 79.1%， respectively. Under high-resolution simulation conditions with a time step of 1 μs， EC-PINN implemented on the GPU platform achieves approximately 6-fold and 10-fold speedups compared with the explicit Euler method and backward Euler method， respectively， demonstrating its application potential for fast time-domain simulation.

To address the issues of poor synchronization coordination and difficulty in stabilizing phase differences in three-motor driven electric-drive axles under vibration synchronization conditions， this paper takes a circumferentially distributed three-motor electric-drive axle driven vibration system as the research object. The kinematic differential equations and the dynamic model of the vibration system are established. The small parameter averaging method is employed to conduct a stability analysis of the three-motor compound synchronization， and the synchronization and stability conditions of the system are derived. Furthermore， by integrating fuzzy adaptive sliding mode control with rotor flux-oriented control， a speed-phase dual closed-loop synchronization tracking controller is designed. MATLAB simulation results demonstrate that the proposed control strategy and method can significantly attenuate phase difference fluctuations among the motors， achieving zero-phase-difference synchronous operation. The system maintains a stable synchronization state under variations in motor structural parameters and internal disturbances. Compared with conventional controlled synchronization approaches， the proposed method reduces the system synchronization settling time by approximately 30% and effectively suppresses chattering induced by mechanical coupling， thereby significantly enhancing the robustness of the distributed drive system under dynamic operating conditions.

To address the susceptibility of visual features to adverse environmental conditions， the limited semantic information provided by LiDAR， and the difficulties of continuously tracking long and slender pipelines during low-altitude inspection， an autonomous navigation method for unmanned aerial vehicles （UAV） based on LiDAR–vision fusion is proposed. A local bird’s-eye-view space is used as a unified multimodal representation to integrate visual semantics， LiDAR geometry， and inertial motion states. A gated state-space model driven by spatiotemporal observation confidence is developed. An adaptive modality-reliability gate is employed to suppress interference from degraded features， while pipeline-topology-guided forward scanning and local neighborhood scanning are introduced to enhance continuous centerline modeling. A local navigation cost function jointly considering pipeline tracking， obstacle avoidance， observation angle， and trajectory smoothness is then constructed. Experimental results show that the proposed method achieves a pipeline-tracking error of 0.17 m， a position root-mean-square error of 0.13 m， an obstacle-region intersection-over-union of 88.2%， an obstacle-avoidance success rate of 96.4%， and a single-frame inference time of 46.8 ms. The results demonstrate that the method can meet the requirements of low-altitude autonomous inspection in typical water-intake pipeline scenarios.

In order to address the challenges brought by heterogeneous environment interconnection and to improve the security of power grid industrial control systems， an automated verification method for security vulnerabilities in power grid industrial control networks under heterogeneous environment interconnection is designed. A source code vulnerability mining graph network， consisting of a decoder， a processing core， and an encoder， is constructed to enable automated mining of security vulnerabilities in power grid industrial control networks. A big data representation framework based on the tensor Tucker decomposition algorithm， consisting of a big data tensor quantization representation module and a big data dimensionality reduction module， is employed to achieve a unified representation of the source code vulnerability data mined. Divide the automated verification process of security vulnerabilities into two key stages： generation of migration paths and driving execution of migration paths. Efficiently generate transition paths between states using the method of constructing transition paths through search， and use a fuzzy testing strategy based on transition path orientation to construct efficient test cases， guiding the program to smoothly reach the predetermined target state. The test results show that the design method can generate different types of vulnerability exploitation code， with a relatively short time consumption， of which the shortest time being only 257 seconds.

To address the mutual coupling challenges among topological relationships， geometric layout， annotation avoidance， and drawing partitioning in the automatic generation of piping isometric （ISO） drawings for process plants， this paper proposes a phased collaborative constraint-solving method. An attribute-directed graph is used to represent connectivity， flow direction， assemblies， and engineering attributes. With isometric directions fixed after discrete mapping， continuous layout and geometric conflict handling are performed through linear programming and incremental separation constraints. For annotation selection， multiple-start search and conflict-edge repair are first used to generate initial solutions. Adaptive Neighborhood Rollback （ANR） then determines active neighborhoods from the residual conflict graph and jointly reoptimizes the candidate combinations within those neighborhoods. Drawing partitioning is incorporated as a feedback step after continuous layout and annotation selection. Experiments using fixed-seed models， ablation settings， and MILP-based optimality verification show that the method maintains topological consistency while coordinating drawing layout， annotation quality， and computational cost across pipelines of varying scales. In particular， ANR is primarily employed for the locally coordinated handling of residual conflicts in quality-oriented offline generation. In dense cross-sheet scenarios， the finite candidate set and limited neighborhood still impose applicability limits， requiring further processing through candidate expansion， neighborhood adjustment， and split-point reselection. The resulting workflow provides a traceable implementation path for automatic generation and quality review of piping isometric drawings.

A small-scale gantry-type five-axis CNC machine tool was designed， and research on structural optimization and performance improvement was carried out around it. First， finite element method was utilized to perform static， modal， and harmonic response analyses on the entire machine， which identified the ram as the critical weak component. Subsequently， topology optimization based on the SIMP （Solid Isotropic Material with Penalization） method was conducted to reconstruct the internal structure and force transmission path of the ram. A response surface model was then established based on parameter sensitivity analysis， and key dimensions were optimized via a Multi-Objective Genetic Algorithm. Simulation results indicate that the optimized ram exhibits a mere 6.16% increase in mass， while achieving a 44.04% reduction in maximum deformation and a 10.05% improvement in the first-order natural frequency. Furthermore， the overall machine tool performance demonstrates significant enhancement， characterized by a 14.97% decrease in maximum deformation and elevated natural frequencies across all modes. Notably， the maximum resonance amplitudes in the X， Y， and Z directions are reduced by 51.39%，39.15%， and 53.21%， respectively. The static and dynamic performance of the entire machine is significantly improved， providing a valuable reference for the structural optimization design of similar small-scale gantry-type CNC equipment.

Given the multi-variety and variable-batch production characteristics of aerospace product assembly lines， the insufficient synergy between resource allocation decisions and process layout planning during dynamic reconfiguration leads to issues such as insufficient capacity and production imbalance. To address these problems， this paper establishes a multi-objective optimization model aimed at minimizing production costs and maximizing production line balancing rate， which is solved using an improved NSGA-II algorithm. The algorithm adopts a three-layer chromosome encoding structure and incorporates logistics-intensity-driven crossover， mutation， and local search strategies to enhance solution quality and search efficiency. Case validation results demonstrate that the proposed model can effectively determine the production line process layout scheme adapted to specific production tasks； compared with NSGA-II， the improved algorithm achieves a better Pareto front and faster convergence.

To suppress far-end crosstalk in coupled microstrip lines of high-speed printed circuit board interconnect systems， this study proposes a composite microstrip-line model that integrates a T-shaped defected structure with a tabbed routing structure. An improved soft actor-critic reinforcement learning algorithm is further employed to optimize the structural parameters of the model. The optimization of the microstrip-line geometry is formulated as a Markov decision process with a continuous action space， in which the closed-loop optimization between the SAC agent and the electromagnetic simulation environment is described by the state， action， reward function， and state transition. Moreover， a dynamic target entropy coefficient adjustment strategy is proposed for the SAC algorithm based on the current optimization state of the microstrip line， thereby enhancing the balance between exploration and exploitation in limited electromagnetic simulation environments. The results show that the improved SAC algorithm achieves significantly faster convergence， with an 8.28% improvement in convergence speed over the conventional SAC algorithm. Compared with the conventional microstrip-line model， the optimized model reduces the average far-end crosstalk from -18.96 dB to -37.41 dB over the broadband range of 0-16 GHz. At the dominant frequency of 3.2 GHz， the crosstalk suppression reaches -59.02 dB. In addition， the absolute peak crosstalk voltage decreases from 41 mV to 4 mV. These results demonstrate that the proposed method can effectively improve the efficiency of the SAC algorithm， obtain optimized structural parameters for the composite microstrip line， and suppress crosstalk interference in high-speed links and improve signal transmission stability and overall system integrity.

Design for Assembly （DFA） is an important approach for improving manufacturability and reducing assembly costs during the early design stage of mechanical products. However， existing DFA evaluation methods mainly rely on manual experience and rule-based inspection， making it difficult to establish cross-domain associations among geometric design information， assembly process knowledge， and manufacturing resource capabilities. Consequently， assembly constraint conflicts cannot be effectively identified and traced at the early design stage. To address the aforementioned issues， this study proposes a knowledge-based reasoning and design optimization method for mechanical-product DFA under assembly constraints. Firstly， an automatic part DFA feature extraction method based on computational geometry is developed. By analyzing the B-Rep geometric topology information of STEP models， structured design features including dimensional parameters， physical properties， symmetry， and stability are extracted automatically， enabling the transformation from three-dimensional models into DFA evaluation data. Secondly， a ternary collaborative ontology model of “part-process-equipment” is established. Through knowledge graph representation， semantic associations among design attributes， assembly processes， and manufacturing resources are constructed， and dynamic mapping of manufacturing resource parameters is achieved. Furthermore， a hybrid reasoning mechanism integrating OWL-based description logic reasoning and Python-based rule reasoning is developed. Combined with a hierarchical 931 scoring evaluation system， the proposed method realizes assembly constraint conflict identification， cause tracing， constraint severity evaluation， and design optimization feedback. A butterfly valve assembly containing 58 components is selected as the experimental case. Two typical constraint problems， including manufacturing equipment capability mismatch and structural design constraints， are analyzed to validate the proposed method. In addition， modular ablation experiments are conducted to investigate the impacts of geometric feature extraction， knowledge modeling， and hybrid reasoning mechanisms on the performance of the method. The experimental results demonstrate that the proposed method can effectively support assembly constraint diagnosis during the early design stage of mechanical products， achieving design problem identification， conflict cause tracing， and optimization decision support. The proposed method provides an effective approach for manufacturing-oriented collaborative design optimization.

To address the lack of language-level compile-time memory safety protection in traditional C/C++ real-time kernels， a preemptive kernel prototype is designed and implemented in Rust for the LoongArch architecture， with the LA464-based Loongson 3A5000 as the reference processor. At the hardware level， LoongArch features， including its register layout and unified exception entry， are exploited to implement context switching through a dedicated context pointer register and a fixed-length instruction sequence. At the safety level， Rust’s ownership and borrow-checking mechanisms are used to confine unsafe operations. Unsafe Rust accounts for 3.1% of the kernel code. The protective effects of Rust mechanisms against eight categories of typical security defects are analyzed with reference to the Common Weakness Enumeration （CWE）. At the scheduling level， a hybrid algorithm is proposed， using static priority as the primary sorting key and dynamic remaining time to deadline as the secondary key. Experiments on the QEMU LA464 model show that the exception-entry path overhead is only weakly related to the number of tasks， whereas the scheduling-decision overhead increases with the task count. In both one-shot task scenarios， the hybrid scheduler achieves a 100% deadline-meeting ratio for high-priority tasks and the lowest average response time among the three strategies； in the heterogeneous conflict scenario， its average response time is 545.80 ms， 17.8% and 20.9% lower than those of fixed priority preemptive scheduling （FPPS） and earliest deadline first （EDF）， respectively. In the periodic-release scenario， all 10 task instances meet their deadlines under both hybrid scheduling and EDF， compared with a 70% overall deadline-meeting ratio under FPPS. The results provide preliminary evidence of the feasibility of the proposed kernel in LoongArch adaptation， Rust safety mechanisms， and hybrid scheduling.