Cold chain transport is the process of keeping goods at low temperatures from one location to another throughout the supply chain， used mainly in food， pharmaceutical， biomedical and other industries to ensure the quality and safety of goods， and the reasonable planning of transport is thus crutial for reducing logistics costs and carbon emissions and improving transport efficiency. This paper establishes a mathematical model for the optimization of cold chain logistics transport paths， taking into account the distance between different customer points， service time window， cargo demand and other factors， in order to achieve the goal of minimizing the total cost of transport， including fixed costs， transport costs， refrigeration costs and so on. The model is then solved using the improved seagull algorithm， which is a bionics-based optimization algorithm that simulates the behavior of seagulls when foraging， and the algorithm has global search capability and adaptivity. During the implementation of the algorithm， this paper considers the iterative parameter setting and convergence control of the algorithm to avoid falling into the local optimal solution and to ensure the reliability and the efficiency of the experimental results. The effectiveness of the cold chain logistics path optimization method based on the improved seagull algorithm is verified through the application of experimental examples and the comparative analysis of different algorithms.

The application scenarios of large-scale equipment data communication in industrial sites requires a data communication solution with low latency， large capacity and high speed. This paper compares mainstream cellular IoT technologies， proposes an industrial site data collection technology solution based on the 5G lite technology RedCap， desigs a data communication terminal based on 5G RedCap technology， and verifies its feasibility through testing. The data communication terminal based on 5G RedCap technology can well meet the needs of industrial site data communication， is in a valuble position to be promoted and applied in the field of IIoT， and will push forward the development and evolution of cellular IoT towards end-network collaboration.

To achieve precise and automated segmentation of the middle section of pork carcasses， an improved YOLOv11-Pose-Based keypoint detection method for rib contour recognition is proposed. Addressing the limitations of existing algorithms in recognition speed， learning efficiency， and detection accuracy， the CBAM attention mechanism's CAM submodule is innovatively replaced with the Bi-level routing mechanism from Biformer， effectively enhancing model efficiency and recognition accuracy. The experimental results demonstrate that the improved model achieves significant performance gains across multiple metrics. For keypoint detection， the mAP（0.5） reaches 0.995， representing a 2.37% improvement over the baseline model， while the mAP（0.5-0.95） increases by 0.97%. The precision of bounding box prediction improves to 0.983， an enhancement of 2.82%. Additionally， the recall rate rises to 96.5%， a 7.8% increase compared to that of the original model. Meanwhile， the training time is reduced by 8.46% without altering the model size， significantly enhancing computational efficiency. The visualization results confirm that the improved model accurately identifies the rib contour of the middle pork carcass section and precisely predicts five key points， providing reliable technical support for subsequent automated segmentation tasks. The findings indicate that this approach achieves high accuracy while improving computational efficiency， offering a novel solution for intelligent pork segmentation technologies.

Tension control is a critical technology in the production line of biaxially stretched film， as it directly influences the quality of the final product. To address the issue of tension fluctuations in thin films caused by tower rotation during roll changes—an effect that compromises finished film quality—this paper introduces a feedforward speed compensation strategy. This approach involves calculating the cumulative speed imparted to the film due to tower rotation during roll transitions and designing a tension control system based on a feedforward-PID framework informed by this calculated relationship. Comparative simulation experiments demonstrate that this system outperforms traditional PID controllers， yielding significant improvements in film production efficiency.

This paper addresses the challenge of inadequate timeliness and accuracy in processing multi-source heterogeneous data within special equipment assembly workshops， which tends to hinder real-time transparent management and control during the assembly manufacturing process of specialized equipment. To tackle this， a method for fusing multi-source heterogeneous data in these workshops is proposed with the approach based on an analysis of the composition and characteristics of operational data in these workshops， and the benefits of Multi-Agent technology is applied to construct a framework for data fusion. The research on the methods involved in data layer fusion and feature layer fusion is also conducted. The feasibility and effectiveness of the proposed method are confirmed through simulation examples， thereby providing reliable， timely， and accurate data support for the intelligent operation and control of special equipment assembly workshops.

Aiming at the mechanical resonance problem in the dual inertia system， a fuzzy sliding mode control method based on unknown input observer is proposed in this paper. This method mainly includes two parts ： the unknown input observer and the fuzzy sliding mode controller. The observer part is designed based on the principle of invariant manifold to observe and compensate the transmitted torque. On this basis， the sliding mode controller is designed to ensure the tracking accuracy of the output signal. In order to eliminate the chattering problem of the traditional sliding mode control itself， a combination of fuzzy control and sliding mode control is proposed at the controller part， which greatly improves the dynamic responsive performance of the system. Finally， the effectiveness of the proposed method is verified by simulation.

To solve the problems of small molding size and low printing efficiency of traditional desktop-level single-nozzle FDM 3D printer， a medium-sized FDM multi-nozzles collaborative 3D printer is designed and built. The printer adopts a Cartesian （XYZ） structure and is equipped with three side-by-side composite printing nozzles， and the materials can be selectively extruded according to the demand. The control system is divided into three parts according to the functions： main motion control module，embedded auxiliary measurement and control module and upper computer software module，while the software and hardware of these three parts are developed.Two printing modes of multi-nozzles synchronous forming and multi-nozzles stackable co-filling are designed and the corresponding path planning algorithms are proposed.After printing verification， compared with single-nozzle printing， the synchronous forming efficiency of the composite multi-nozzles printer is increased by 3 times， whereas the stackable co-filling printing time is reduced by 41%. The printing efficiency is significantly improved under the premise of ensuring the printing quality.

Heat exchangers play a crucial role in improving the energy efficiency of industrial processes， reducing fuel consumption， and decreasing greenhouse gas emissions. This paper addresses the innovative design problem of heat exchangers with numerous parameters， variable structures and complex medium flow characteristics by proposing a generative design method for spiral tube heat exchangers. Firstly， it analyzes the design principles， the structural advantages， and the performance characteristics of spiral tube heat exchangers， introduces the application process of the generative design method， the design optimization logic， and the automated parametric model generation method. Then， through computational fluid dynamics simulation， it evaluates the thermal transfer efficiency and fluid dynamics performance advantages of the spiral tube heat exchangers. Finally， through structural mechanics simulation， it assesses the risk resistance performance advantages of the spiral tube structure under various operating conditions. The proposed generative design method achieves rapid optimization iteration of design solutions and rapid generation of heat exchanger models， providing the possibility for rapid exploration and design of high-performance spiral tube heat exchangers.

To ensure the safety and stability of the power transmission systems and achieve comprehensive inspection and maintenance of newly constructed power transmission towers， this paper proposes a centrally symmetric quadrupedal humanoid climbing robot designed for existing foot pegs used by maintenance workers to climb towers. Each limb is configured with 3-1-2 arrangement. At the end of each limb， a large-tolerance， semi-enclosed hook-type gripping tool is designed specifically for the foot pegs. This tool features high tolerance， eliminating the need for precise end-effector positioning and enabling rapid engagement with the foot pegs. Humanoid climbing gait planning method is developed， facilitating the robot's full-range climbing of the power transmission tower by quickly hooking and gripping the foot pegs using the hook-type tool. Targeting a 40-meter-high self-supporting transmission tower， the robot's full-range climbing dynamics model and simulations were completed. Simulation results demonstrate that the proposed robot configuration can achieve humanoid full-range climbing of the tower， with a climbing time from the base to the top of less than 30 minutes， matching the efficiency of maintenance personnel. This provides a feasible solution for robotic maintenance applications in power transmission towers.

Detection on the surface defect of Pouch Cells is a key procedure of the production process. Aiming at the problems of low detection accuracy and difficult imaging of large-sized batteries in the existing detection methods， a detection method based on photometric stereo imaging and deep learning was proposed. Firstly， a Multi-Source Time-Sharing Exposure Imaging System （MSTIS） was established by combining photometric stereo and line scan camera imaging technology. After obtaining the surface images of batteries under different light sources through time-sharing exposure， photometric stereo calculation was conducted to obtain the curvature map with 3D information. Then， to solve the problem of missed detection of minor target and multi-scale defects， the YOLOv8 algorithm was improved. An edge information enhancement module （EIEM） was developed using a dual-channel convolution structure， which incorporated Sobel convolution and conventional convolution to improve feature edge extraction capabilities. The semantic and detail information fusion method （SDI） was integrated with the bidirectional feature pyramid module to boost the recognition accuracy of tiny defects. A lightweight shared convolution detection head was also implemented to reduce the algorithm's computational load.The experimental results show that the average detection accuracy of this method reaches 94.2% and the detection speed reaches 116 FPS， which can effectively detect the surface defects of pouch cells.

Accurate force/position control in the presence of modeling uncertainty in robots is a challenging task. This article is based on improved integral sliding mode control for hybrid visual & force control. To perform visual servoing， an optimized visual feature is adopted to avoid the ill-condition visual Jacobian matrix. An improved super-twisting algorithm is proposed based on time delay estimation for integral sliding mode control， and an improved sliding surface and convergence law is used for the positive definite part of integral sliding mode control. Correspondingly， visual control method and force control method are proposed to apply to the hybrid visual & force control framework. To avoid the impact of noise from the force control part on the control output， a visual admittance framework is used for hybrid control. To optimize the solidification impedance characteristics caused by fixed admittance parameters， a fuzzy adaptive admittance framework is proposed to inherit the advantages of admittance while adaptively adjusting parameters in real-time. Finally， a 6-degree-of-freedom deviation model is used for simulation to verify that the proposed scheme can accurately track visual trajectories and expected forces under relatively small chatting. The performance of the three with different focuses is compared based on direct visual perception hybrid control， visual admittance control framework， and fuzzy adaptive admittance framework. Under the deviation model， the proposed i-CISMC algorithm is validated by surface force tracking to have better tracking accuracy while ensuring smaller control chattering； the proposed algorithm is combined with the fuzzy adaptive admittance framework， and the results of tracking under noise have proved that the framework inherites the suppleness of the admittance framework to the disturbance of the force loop， and at the same time， can adaptively change the parameters of the admittance to obtain better tracking speed and tracking accuracy.

To address the challenges of frequent bearing failures under complex working conditions， as well as the low real-time performance and strong dependence on manual feature extraction in traditional diagnostic methods， this paper proposes a bearing fault diagnosis method based on a deep learning model combining a Multi-Scale Convolutional Neural Network （MSCNN） and Long Short-Term Memory （LSTM）， and develops an intelligent bearing health management system. The system adopts an end-to-end diagnostic workflow， directly taking raw time-domain vibration signals as input. It extracts hierarchical local features across different frequency domains through MSCNN， and captures the temporal evolution of fault characteristics using LSTM， thereby achieving high-accuracy automated fault classification. To enhance the interpretability of diagnostic results and support intelligent maintenance decisions， the system integrates the Chinese large language model iFLYTEK Spark， which generates natural language diagnostic reports and maintenance suggestions through standardized prompts. The system is deployed on a domestically developed Phytium quad-core processor platform， ensuring full autonomy and reliability of both hardware and software components for industrial applications. Experimental results show that the proposed system achieves an average classification accuracy of 98.46% on the CWRU bearing dataset， and 96.73% on the AITHE bearing fault dataset， demonstrating strong robustness and cross-dataset generalization under complex and noisy conditions. With real-time visualization of diagnostic results and maintenance recommendations through a human-machine interface （HMI）， this system provides a reliable and intelligent solution for equipment health management and predictive maintenance.

To address the issues of low spatial utilization and Inbound and outbound efficiency in cargo location allocation for mechanical processing warehousing scenarios， a hybrid model （GA-XGBoost） integrating Genetic Algorithm （GA） and XGBoost parameter optimization is proposed. By constructing a two-layer encoding mechanism for feature selection and hyperparameter collaborative optimization and combining an improved greedy algorithm with dynamic priority adjustment， a multi-constraint decision model is established with optimization objectives of spatial utilization， inventory time， and prediction accuracy. The experiments based on warehousing data of 500 cargo locations and 1，200 types of goods show that the average in/out time is reduced to 17.9 minutes， representing an 18.7% efficiency improvement； the mean squared error of prediction is reduced to 0.012， and the number of convergence generations is decreased by 19.4%. This method effectively balances multi-objective constraint relationships and provides a dynamic cargo location allocation scheme for intelligent warehousing systems that coordinates high-density storage and efficient operations.

To address issues such as low detection accuracy， slow speed， missed and false detections， and large model parameter sizes in complex scenarios from a UAV perspective， this paper proposes an improved RBGE-YOLO algorithm model. Firstly， RFAConv is introduced in the backbone network to replace the original Conv， enhancing the model's ability to extract and fuse image features. Secondly， the neck network is reconstructed using BiFPN-GLSA to improve feature fusion and spatial feature utilization efficiency. Thirdly， a dual-layer small target detection structure is designed to strengthen the feature information of small targets. Finally， the Inner-EIoU loss function is utilized to address the limitations of IoU. Experiments on the VisDrone2019 dataset show that RBGE-YOLO improves Precision， Recall， mAP@0.5， and mAP@0.5：0.95 by 4.7%， 2%， 3.6%， and 2.5%， respectively， compared to the original YOLOv8s， while reducing the number of parameters by 16.4%. This achieves model lightweighting while significantly enhancing detection performance.

With the rapid development of FDM 3D printing technology， the demand for printing efficiency and molding accuracy is increasing cross various industries. This paper proposes a motion system control strategy based on a combination of BP neural network and PID control algorithm to address the issue of insufficient forming accuracy in Delta structured FDM 3D printers. By introducing BP neural network into the motion control system of the printing device， dynamic adaptive adjustment of control parameters has been achieved， effectively improving the stability and accuracy of the printing process. The experimental results show that this control strategy significantly reduces printing errors and improves the surface quality and dimensional accuracy of the molded parts. This study provides a new solution for improving the accuracy of Delta structured FDM 3D printers， which is of important theoretical significance and practical application value for promoting the development of 3D printing technology.

In the field of modern logistics warehousing， carton identification is crucial for inventory management and logistics automation. Aiming at the shortcomings of traditional methods and some existing automation schemes in carton detection tasks， a new carton detection method based on improved YOLOv8 network is proposed. Firstly， an Adaptive Batch Normalization（ADBN） mechanism is proposed and introduced to the YOLOv8 backbone network， enhancing the feature extraction ability. The C2f-Faster-CGLU mechanism combining FasterBlock and Convolutional Gated Linear Unit （CGLU） is introduced into the YOLOv8 detection header， which improves the computational efficiency. In addition， a new boundary frame similarity comparison index based on the minimum point distance （MPDIoU） is introduced， which can evaluate the similarity between prediction and real frame more accurately. Finally， the improved network model is applied to Rectagular Stacked Carton Dataset（RSCD）， Online Stacked Carton Dataset（OSCD） and Live Stacked Carton Dataset（LSCD）. Compared with that of the original model， the mAP of the improved model is increased by 1.6%， and the recall rate is increased by 1.3%. The improved model has also improved performance compared with other mainstream detection algorithms，providing more accurate and efficient technical support for the object detection of modern logistics and warehousing industry.

To address the problems of scattered knowledge systems in ship assembly and ineffective mining and utilization of massive process data， this paper proposes an automatic knowledge graph construction technology for the shipbuilding domain based on large language models. This method uses large language models to convert unstructured and semi-structured ship data into structured data to build a ship process corpus. It models ship ontology knowledge structure with the assistance of large language models， designs an instruction prompting framework for ship assembly domain， and achieves efficient entity-relationship extraction， to complete the automatic construction of knowledge graphs. Additionally， the method uses triple sets constructed by general large language model instruction prompts as fine-tuning training sets to further fine-tune specialized small language models， ensuring the security of specific private ship data while reducing computational resources. Experimental results show that this method outperforms traditional baseline models in key metrics such as accuracy， providing a new technical approach for knowledge management and intelligent upgrading in the shipbuilding domain.

Most amphibious biomimetic robots encounter problems such as insufficient motion ability， poor environmental adaptability and low simulation rate. This article adopts a novel central pattern generator （CPG） with dual neuron mutual inhibition as its main controller based on the basic rhythmic gait of salamanders， and ensures the phase coupling relationship between adjacent CPG units by adjusting the excitation suppression parameters between each neuron. Based on this， a salamander robot spinal cord like control neural network is established. The neural network consists of two layers：Interneuron and Motor neuron. The Interneuron layer generates rhythmic signals， which are then integrated by the Motor neuron layer before outputing to the joint muscle model to drive the robotic movement The performance of spinal cord control network was simulated and analyzed by combining Simulink and Webots. The simulation results show that the amphibious salamander biomimetic robot can effectively achieve rhythmic gait such as swimming and land crawling. The neural network for motion control of the salamander robot designed in this paper is feasible and effective.

Bird activities have threatened the stable operation of electrical equipment in the substations. To address the current equipment's low detection accuracy and poor bird-repelling effect， this article proposes a method that combines an audio-visual perception module， an acoustic-optic bird repelling module and the remote control. The module-in-one intelligent bird detection and repelling robot system uses the time difference of arrival （TDOA） algorithm based on particle swarm optimization to realize the sound source positioning of the microphone array and applies the improved YOLOv5 and DeepSORT algorithms to achieve precise positioning and tracking of bird targets. Finally， the targeted bird repelling is achieved based on acoustic-optic equipment. The experimental results show that the distance estimation accuracy of the sound source reaches 96.42%， the phase estimation error is less than 1.1°， and the visual recognition accuracy reaches 89.7%. The fusion of audio-visual modal data effectively improves bird detection accuracy. The problem of blind spots in the visual field existing in the bird detection process can be solved faster and more accurately to complete the task of detecting and driving birds.

Aiming at the problem that traditional convolutional neural networks cannot effectively extract global features and some deep learning models are more complex， this paper proposes a lightweight gearbox fault diagnosis method based on multi-scale convolutional neural networks and feature fusion ViT. Firstly， a multi-scale feature extraction module is constructed， which captures the feature information of the data from multiple scales by multi-scale convolutional neural network using different scale convolution kernels， and fully exploits the local features of the input information. Then， the feature fusion ViT module is designed， which utilises an improved multi-attention mechanism to capture the global features of the fault information， and further constructs the D-MLP to reduce the number of parameters in the model using depth-separable convolution. Finally， the experimental validation is given using gearbox data from Southeast University， and the results show that， compared with the comparison methods， the proposed method has high fault diagnosis accuracy and good generalization ability under complex conditions such as variable operating conditions and variable noise.

To improve the quality and efficiency of assembly and ensure the consistency in product quality， the assembly actions of workers are recognized and monitored. This paper proposes a fusion assembly action recognition network with Longformer's spatiotemporal separated attention. Through the spatiotemporal separated attention structure， Longformer attention encoder and Transformer attention encoder are used separately to extract appearance and motion features of the video， effectively integrating spatiotemporal information in long video sequences while reducing the computational complexity and network parameters. Experimental results on an assembly action dataset show that our approach outperforms the convolution-based SlowFast network in extracting global video features， achieving a 2.44% improvement in Top-1 accuracy. Compared to the Transformer-based TimeSformer network， Top-1 accuracy is improved by 0.45%， and parameters are reduced by 65.9%， while enabling effective recognition of worker assembly actions with longer video sequences.

To solve the dynamic job shop scheduling problem in scenarios with variable job and machine quantities， a solution approach called Dense-D3QN， combining DenseNet， a densely connected convolutional network， with Dueling Double Deep Q-Learning （D3QN） is proposed. The disjunctive graph model is utilized to construct a single-objective job shop scheduling model aiming to minimize the maximum processing time， representing the scheduling state in the form of multi-dimensional matrices while designing a dense-sparse reward function. To validate the effectiveness of the proposed algorithm， both public benchmarks and real data are used to construct common and actual scheduling environments. The Dense-D3QN model is trained and tested in the common environment. In the actual environment， the Dense-D3QN model is trained and tested in both static and dynamic settings. The experimental results demonstrate that the Dense-D3QN model is more capable of handling dynamic job shop scheduling problems with variable scales.

As an important component of intelligent manufacturing and intelligent operation and maintenance systems， industrial inspection robots are playing a key role in various complex industrial scenarios. With the continuous progress of deep learning， multi-sensor fusion， and autonomous navigation technologies， industrial inspection robots have significantly been improved in terms of accuracy， efficiency， and adaptability. This article systematically reviews the concept， key technologies， and typical applications of industrial inspection robots， and focuses on analyzing the research status of core technologies such as perception and recognition， autonomous positioning and navigation， advanced control， and intelligent decision-making. It also assesses the maturity and industrialization progress of current technologies by combining practical applications in fields such as power， workshops， and special environments. Despite significant achievements in this field， challenges still exist in perception accuracy， dynamic environment adaptability， and task execution intelligence. The development of key technologies is expected to continue in the directions of multi-source data fusion， autonomous learning， and collaborative operation. The article aims to provide a systematic reference and guidance for future research and industrial development of industrial inspection robot technology.

To improve the path planning efficiency of mobile robots and solve the shortcomings of Gray Wolf Optimization algorithm such as low convergence efficiency and suscepitability to falling into local optimal in path planning obstacle avoidance， a Particle Swarm Optimization improved Gray Wolf Optimization algorithm is proposed. Firstly， Tent chaotic mapping is first added to initialize the initial population to increase the diversity of the population and thus improve the convergence speed. Secondly， the nonlinear convergence factor improvement strategy is added to improve the efficiency of global search while reducing the local optimal solution. And then， the particle swarm location updating strategy is applied to the gray wolf population location updating to enhance the autonomous searching ability of individual gray wolves. Finally， compared with PSO and GWO， the improved algorithm has superior convergence performance and optimization accuracy. Simulation results show that the improved algorithm is superior to other algorithms in average path length， iteration times and search time.

Irregularly shaped objects often appear inside the pipelines of nuclear power systems， and soft grippers with passive deformation ability are hence needed for grabbing operations. While the load capacity of the fixed stiffness soft gripper is usually low， the variable stiffness is an important way to improve the load capacity of the soft gripper. On the other hand， the force perception is very important for grasping fragile objects， but rigid force perception is difficult to integrate with soft grippers. Therefore， it is crucial to develop a flexible force perception function module that can be integrated with the soft gripper. Based on the characteristics of the tendon-air hybrid drive structure and the coordinated working relationship of the muscles on the upper and lower sides of the bionic human fingers， this paper proposes a design idea of a soft gripper that uses antagonistic action to change the stiffness of the soft fingers and integrates a force perception function. The designed soft gripper has three pneumatic chamber fingers， each of which is prepared by the distributed casting silicone rubber process， and the deformation process of a single pneumatic chamber finger is simulated using simulation software. To achieve contact force perception during grasping， a force sensing system based on Hall chip magnetic field intensity sensing is integrated at the tip of the soft finger， and the performance of the sensor is calibrated and tested. Through experiments， the adaptive grasping ability of the soft gripper for different target objects and the load capacity of the tendon-air antagonistic drive are tested. The experiments show that the tendon-air antagonistic drive can effectively improve the load capacity of the soft finger， and the perception module can achieve real-time measurement of the contact force of the grasping target.

In the industrial software ecosystem， Computer-Aided Design（CAD） interfaces play a pivotal role. This study outlines the composition and collaborative mechanisms of the industrial software ecosystem， reviews the evolutionary trajectory of CAD interface technologies， and summarizes their core roles within the ecosystem from the perspectives of data transmission， functional integration， and innovation-driven development.Building on this foundation， an in-depth analysis of the application bottlenecks and challenges faced by CAD interfaces is conducted， including data interface standards， the depth of system integration， and the convergence with emerging technologies. Furthermore，future development trends for CAD interfaces are explored， emphasizing key directions such as data standardization and semantic enrichment， multi-user collaborative design with real-time interaction， and the deep integration of artificial intelligence technologies. This work aims to provide theoretical insights and practical guidance for the research and application of CAD interfaces within the industrial software ecosystem.

By analyzing the existing gangue sorting manipulator claw and its use， it is concluded that the existing claw has a large weight， is susceptive to wear and tear as well as higher cost of the overall replacement. A method is hence adopted to install replaceable wear-resistant shims and to select lighter quality materials for improvement. The finger force analysis is made before and after the improvement through the Ansys Workbench， and the improved finger effect proves to be better， verifying the feasibility of the installation of replaceable wear-resistant shims， while determining the replaceable wear-resistant shims material being 20CrMnSi， and finger base material being TC4. The fatigue life analysis is made for the finger before and after the improvement using fatigue analysis tools， and the conclusion is drawn that the fatigue life of improved finger matrix is longer， and the replaceable wear-resistant shims begin to fail after being used 4.3794×105 times， and are therefore needed to be replaced after about two months of use.

The advancement of Industry 4.0 necessitates the deployment of intelligent， low-cost robotic systems on edge devices. However， the high computational complexity of Deep Reinforcement Learning （RL） algorithms presents a major obstacle to their implementation on resource-constrained platforms such as the Raspberry Pi. To overcome this challenge， this paper introduces a lightweight RL framework tailored for industrial robot sorting tasks. The core contributions are threefold： First， we propose a joint compression method combining Gradient Sensitivity-guided structured Pruning （GS-Pruning） with hierarchical quantization， which reduces model size by over 90% and achieves real-time inference below 35 ms on a Raspberry Pi while preserving policy accuracy. Second， we design a Dynamic Weight Adaptive Reward function （DWAR） that balances sorting efficiency， motion stability， and energy consumption， successfully suppressing robotic arm jitter and cutting average energy use by 18.1%. Third， we construct an end-to-end deployment system， RPi-EdgeRL， featuring a multi-threaded pipeline and a safety watchdog to guarantee stable and efficient autonomous operation. Experiments conducted on a FR3 collaborative robot validate our framework， achieving a 93.5% success rate in complex sorting tasks and confirming the feasibility and superiority of this low-cost， high-efficiency solution for real-world industrial applications.

Industrial robots， with their high flexibility and large working range， have gradually become another important processing equipment besides CNC machine tools in national strategic fields such as aerospace and maritime industry in China. The dynamic characteristics of the robot end are dominated by its joints. To improve the dynamic performance of the robot， it is necessary to start with its weak joints， make improvements and innovations on the basis of the existing series form， and explore new high-stiffness driving methods and robot configurations. A 2-RPR robot for milling large propellers is proposed， which includes a six-axis robot main body and a double electric cylinder branch chain. The translation of the electric cylinder drives the rotation of the robot joint， thereby improving the stiffness of the whole robot. In order to meet the processing space requirements of large propellers， the length parameters of each link of the robot are optimized based on the genetic algorithm to realize the optimization of the working space of the whole robot， so that the working space meets the processing range of a single blade and has the maximum utilization rate.

To improve the accuracy of coal and gangue identification in the process of top coal caving， a multi-channel information fusion method for coal gangue identification combining variational mode decomposition （VMD）， principal component analysis （PCA） and convolutional neural network （CNN） is proposed. Firstly， the coal gangue slip test bench of caving hydraulic support is built， and the vibration data of coal gangue in multiple sensor channels are collected. Secondly， the collected data are decomposed into intrinsic mode function （IMF） using VMD， and the time domain and frequency domain features of each IMF component are calculated respectively. Then， the features are reduced by PCA to obtain the feature vector of coal gangue vibration signal after dimension reduction and construct the feature data set of coal gangue vibration signal. Afterwards， the CNN model is used to train different channel data respectively. Finally， the weighted average method is used for multi-channel information fusion to conduct comprehensive evaluation and identification. The results show that the method has a high accuracy of coal gangue identification， the accuracy rate can reach 97.15 % after information fusion， and the coal and gangue identification effect is good.

The quality of screw tightening greatly affects the safety of mechanical products， whereas traditional diagnosis approaches are time-consuming and imprecise， and the implementation of effective fault diagnosis， therefore， bears significant engineering value. In this paper， an innovative method of fault diagnosis for screw-tightening based on LSTM and Expert knowledge is proposed. Firstly， tightening process curve under specific failure mode was studied and several expert knowledge rules were established. Secondly， a data pre-processing algorithm was established based on the characteristics of sequential data such as noise clipping， stage segmentation， fitting and sampling to improve the quality of data. After that， the feature vector obtained through LSTM was used as the input of the expert knowledge model to obtain the expert knowledge vector， and the two vectors were combined as the input of the classifier. Finally， compared with SVM and LSTM， the results show that the method has higher diagnostic accuracy in multiple failure modes.

Based on human-machine coupling modeling and biomechanical analysis， a shoulder-elbow rehabilitation assistive device featuring 5 degrees-of-freedom （DoF） rotational joints and 3-DoF sliding adjustments was developed. Motion capture experiments were conducted to obtain personalized scaled musculoskeletal models and reproduce upper limb rehabilitation movements through inverse kinematics. Utilizing Hill-type muscle models and the Computed Muscle Control （CMC） algorithm， the study analyzed muscle forces and energy consumption during rehabilitation training. Results demonstrated significant reductions in muscle forces for primary movers under assistive support： the long head of biceps brachii showed a 51.34% average force reduction， while the lateral head of triceps brachii exhibited 49.05% decrease. Energy consumption decreased by 30.74% and 36.56% in the long and short heads of biceps brachii respectively， with peak reductions exceeding 40%， indicating sustained unloading effects during elbow motion. Secondary muscles including the posterior deltoid and medial head of triceps brachii maintained moderate 10% reductions， balancing unloading requirements with joint stability to prevent over-intervention. The analysis confirms that the rehabilitation assistive device effectively reduces muscular burden and energy expenditure during training， mitigates muscle overload risks， and provides efficient assistance for patient rehabilitation.

Aiming at the issues of both the high accuracy requirements for identification and positioning of the refractory brick in the automated gripping of coke oven masonry robot and the limited computing resources of masonry equipments on site， a lightweight algorithm is proposed for identifying and positioning the refractory bricks. First， to target misdetection or omission due to the tight arrangement of dry-laid refractory bricks， the self-made refractory brick dataset is enhanced by a multi-boundary feature stochastic synthesis method. The global attention module （GAM） and the loss function Wise-IoUv3 are embedded in the YOLOv8 model to improve the feature extraction capability and the anchor frame quality. Second， to address the problem of limited computing resources of the equipment， the model is pruned by a parameter magnitude-based pruning algorithm to reduce the computational and parametric quantities of the model； Then， based on 2D identifying information， the positioning depth information obtained by a lightweight depth information threshold segmentation algorithm based on point cloud data is fused to accomplish accurate identification and positioning of refractory bricks； Finally， the test is validated by building an experimental platform for coke oven masonry robots. The experimental results show that the recognition mAP@0.5 of the algorithm in this paper reaches 98.83%， while the Params and the FLOPs reach 0.211 M and 0.856 G， respectively， which is an increase of 0.21 percentage points compared to the YOLOv8 model mAP@0.5， and a decrease of 92.99% and 89.56% for Params and FLOPs， respectively. The average relative measurement error is 0.28% when measuring the depth information from 800 mm to 1000 mm. The algorithm in this paper can ensure the recognition and localization accuracy while reducing the requirements of the equipment on computational resources， providing technical support for coke oven masonry robots to realize the automation applications.

To realize the effective continuous target detection and tracking of the insulator flushing robot in the occlusion environment， a detection and tracking method based on computer vision in the occlusion environment is proposed. Firstly， the attention mechanism is added to the YOLOv5 detection algorithm to enhance the recognition accuracy of the detection algorithm， the scale filter in the DSST algorithm is then combined with the KCF tracking algorithm to make the KCF scale adaptive. Thereafter， the Multi-PROSAC-ORB occlusion recognition algorithm is constructed to realize the occlusion recognition. Finally， the above three algorithms are fused and an occlusion judgment condition is proposed to ensure the continuous and stable recognition and real-time performance of the target in the case of occlusion. The experimental results show that the proposed method can effectively avoid the low accuracy of target tracking in the occlusion environment while ensuring the real-time performance， and the target tracking accuracy is increased by 10.9% and the tracking success rate is increased by 13.6% compared with the target tracking accuracy in the unocclusion recognition， which has high accuracy and real-time performance.

In response to the challenges posed by the high annotation costs and the fluctuating annotation accuracy associated with the traditional manual "capture-annotate" approach used in the unordered sorting logistics process of smart factories， an automated data generation method is proposed. This paper focuses on the algorithm for automatic generation of image datasets， establishing a simulation environment based on the Pybullet open-source physics simulation engine. Through the configuration of virtual environment rendering parameters and processing of segmented images captured by virtual cameras， the goal of generating high-quality image datasets and annotation files is achieved. The dataset generated from the virtual environment is employed to train the YOLO v5m target detection network， and the trained model is subsequently applied to real scene images captured by the Kinect v2 camera. The experimental results demonstrate a 92.1% detection rate for objects， 98.2% correct classification rate for two categories of objects， and an average recognition accuracy of 97.6%. The time taken to generate a single image and label file is less than 1 second， making this approach more suitable for efficiently generating standardized image datasets compared to manual methods.

At present， close range manual handles are still used at many bridge crane workplaces. To address the potential safety risks faced by operators， a remote path planning bridge crane experimental device is designed to improve the stability and gripping efficiency of the bridge crane and determine the mechanical structure scheme of the device； By studying the working characteristics of the experimental platform in three-dimensional space， the three-dimensional path planning of the device is designed； the ant colony algorithm is improved by storing pheromones on path nodes， and a search method that combines layer by layer advancement with grid plane method is used， optimizing path nodes using pruning algorithm， and updating pheromones using a combination of global and local path planning. Through the above improvement strategies， the improved ant colony algorithm obtained through simulation in MATLAB software has 3 fewer iterations， 46 fewer inflection points， 22.6874 s shorter algorithm time， and 4.4043 units shorter shortest path compared to the traditional ant colony algorithm. Finally， the operating system of the experimental platform is designed， and an experimental prototype is built. The results show that the device meets the actual work requirements， verifying the feasibility and effectiveness of improving ant colony algorithm.

At present， the equipment related to curved surface conformal antenna printing is mainly multi-axis direct writing， which tends to have low precision， low efficiency and high production cost. It is difficult to achieve the mass production due to the complexity of the path in the printing process brought about by the difficulty in speed matching，and difficult as well for the ordinary two-dimensional array inkjet printing equipment to print a precise pattern on the spatial surface in order to achieve the consistency of the anisotropic direction. According to the actual production requirements and combined with the motion characteristics of multi-axis inkjet forming device， this paper establishes a motion model for the curved surface inkjet forming droplet， designs a detection method for the image processing and key point on the surface of unidirectional curvature change， and builds a detection and control system for a three-dimensional surface array inkjet printing position accuracy. The experimental results show that the key point detection accuracy of this system is perfect in terms of effectively reducing the droplet position error caused by the factors such as vibration， feed speed instability， jetting distance changes， printing frequency changes of the multi-joint robots. Ultimately， the method in this paper is applied to the formation of a double-ring resonator unit antenna pattern on the cylindrical surface.

To address the need for both independent control of Continuous Damping Control （CDC） dampers and coordinated control of the entire vehicle semi-active suspension system， while also improving upon the issues present in traditional semi-active suspension controller software design such as challenges in meeting real-time requirements and low CPU utilization in bare-metal development environments， this study proposes an innovative approach. Initially， the study establishes separate models for the seven-degree-of-freedom semi-active suspension system and the forward-inverse models of CDC dampers. Building upon the skyhook control strategy， the study integrates a vehicle-coordinated parallel fuzzy controller based on the Mamdani fuzzy control method. Subsequently， by transplanting the FreeRTOS-SMP multicore real-time operating system and utilizing the Infineon AURIX series 32-bit triple-core microcontroller TC275 as the main control chip， the study designs the software and hardware system for the CDC damper control unit. Furthermore， the study conducts task scheduling verification of the multicore real-time operating system and validates the effectiveness of the designed control unit and proposed strategy through hardware-in-the-loop testing using typical random road surfaces to demonstrate the improvement in overall ride comfort of the vehicle.

Facing the requirements for diverse starting and ending positions as well as velocity in the field of robotic motion control， this paper proposes a novel universal S-curve velocity planning algorithm aiming at adapting to arbitrarily specified starting and ending positions and velocity conditions. The paper first introduces the velocity-to-velocity S-curve velocity planning algorithm， then elaborates on the general 7-segment S-curve velocity planning algorithm. Based on this， a more versatile S-curve velocity planning algorithm is proposed. For different input parameters， this paper classifies the S-curve （s-t curve） into ten types and provides detailed segmented planning strategies for each type. Through simulation tests， it is verified that the algorithm not only excels in efficiency but also demonstrates outstanding performance in the smoothness and positional accuracy of the planned curve. Furthermore， tests conducted on actual robotic platforms further confirm that this algorithm can effectively reduce shocks and vibrations during robotic operations， significantly enhancing the operational performance of robots while demonstrating good practicality and broad application prospects.

The dexterous multi-fingered robotic hand， serving as a key end-effector， is pivotal for enabling robots to perform fine-grained grasping and compliant manipulation. Its advancement holds significant importance for promoting automation in manufacturing， enhancing the intelligence of service robots， and expanding applications in specialized environments. Focusing on humanoid multi-fingered dexterous hand technologies， this paper systematically reviews the current state-of-the-art and future trends. It begins by elucidating the fundamental concepts， system architecture， and typical characteristics of dexterous hands. This is followed by a comprehensive of research achievements from domestic and international teams and commercially available mainstream multi-fingered dexterous hand products， covering various degrees-of-freedom designs and their respective hardware and software implementations. Key technologies， including core hardware components， multi-modal sensory fusion， and control strategies， are critically analyzed. The paper subsequently summarizes practical applications across domains such as industrial assembly， daily life assistance， and operations in extreme environments. Current challenges， particularly in reliability， multi-modal coordination， generalization capability， human-robot safety， and integration and application， are identified. Finally， future research directions are prospected from multiple perspectives， including standard establishment， novel mechanical structures， advanced multi-modal perception and fusion， bionic evolution， and embodied intelligence， aiming to provide valuable insights for in-depth research and groundbreaking applications of dexterous hands.