Abstract:Based on the YOLOv8n model, this paper proposes a improved hazard detection model CML-YOLO for power tower components. It aims to solve the problems of low accuracy, large number of parameters, high computational complexity and large model weight of multi-scale power tower components hazard detection model under complex background. It is mainly used for the detection of targets such as damaged insulators, rusted dampers and bird nests. Firstly, the C2f-HEFE module is designed to enhance the ability to distinguish between background and target by enhancing the edge information. Secondly, the MSFFPN module is designed, and the multi-scale feature fusion is used to enhance the adaptability of the model to multi-scale targets. Finally, the lightweight LSBDH module is designed to reduce the number of parameters and calculation amount of the model. Experimental results show that compared with the baseline model YOLOv8n, the mean average precision of CML-YOLO is improved by 4.4%, and the number of parameters, calculation amount and model weight are reduced by 33.9%, 20.9% and 26.4% respectively. This model improves detection performance while maintaining its lightweight characteristics, achieving a good balance between model detection accuracy and model weight.

Abstract:To address the issues of local optima, path oscillations, and goal unreachability in traditional artificial potential field (APF) methods for multivehicle collaborative formation obstacle avoidance, this paper proposes an improved APF approach. By implementing four key enhancements—defining the minimum potential energy for gravitational fields, incorporating Euclidean distance into repulsive fields, constructing road boundary repulsive potential fields, and establishing nonlinear formation stabilization force potential fields—the dynamic equilibrium mechanism between attraction and repulsion forces is optimized, thereby improving formation obstacle avoidance capability and driving stability. Numerical simulations demonstrate that in triangular formation obstacle avoidance scenarios, the improved algorithm achieves a 37.7% reduction in arrival time (22.3 s), 23.2% shorter total path length (55.7 m), and 61.5% faster formation recovery (2.5 s), while eliminating local optima and reducing goal unreachability rate from 25% to 2%. Physical prototype experiments further validate the algorithm′s robustness in dynamic environments, showing rapid restoration of triangular formations post-obstacle avoidance. This method provides an efficient and stable solution for multi-vehicle collaborative obstacle avoidance, demonstrating significant application value for intelligent transportation systems.

Abstract:Aiming at the problems of insufficient range, high maintenance cost and insufficient inspection frequency of online monitoring equipment in transmission line inspection, this paper designs a U-type triple-coil wireless power supply system using a U-type triple-coil wireless power supply system and at the same time proposes a parameter design optimization method based on a genetic algorithm in order to realize wireless power supply of online monitoring equipment for 110 kV medium and long-distance transmission towers. On this basis, the circuit model of the U-type triple-coil wireless power transmission system is firstly established, and the relational equations between the output power, transmission efficiency and the mutual inductance of the coil, coupling coefficient, and load impedance are deduced, and then the genetic algorithm is used to search for the optimization of the parameters and obtain the optimal solution and its corresponding system parameter values. Finally, an experimental platform is built according to the simulation data, and the experimental results show that the wireless power supply system for on-line monitoring equipment of U-type triple-coil transmission towers achieves an output power of 81.19 W at an operating frequency of 380 kHz and a transmission distance of 1.2 m, which verifies that it can meet the power supply requirements of on-line monitoring equipment.

Abstract:In order to solve the problems of low accuracy and poor effect of tomato leaf disease detection in natural environment, a tomato leaf disease detection model based on optimized YOLOv8 was proposed, namely GDDL-YOLOv8n. In this model, the original backbone network is improved by using GhostHGNetV2, C2f-DWR-DRB is used to improve the neck network feature fusion, and the Lightweight Shared-Convolutional detection head (LSCG) is innovatively introduced. The lightweight and high-precision detection effect of the model has been successfully realized. Experimental results show that the GDDL-YOLOv8n model decreases by 49.13% in the number of parameters, 37.04% in the amount of computation, and 46.67% in the memory occupation of the model, while maintaining the high-precision detection performance, with the mAP@0.5 reaching 98.4% and the mAP@0.5-0.95 reaching 92.3%. In addition, a user-friendly interface system based on PyQt5 was developed, which supports image and video detection and camera real-time tracking and recognition functions, and the intelligent management of agriculture and the identification technology of tomato leaf disease have been significantly enhanced, and the model is lighter, which greatly promotes the application of these technologies in actual production.

Abstract:Flow cytometry is a high-throughput detection technique widely used in life science research and clinical diagnostics. However, conventional flow cytometers exhibit suboptimal performance when handling complex data dimensions and severe noise interference. To enhance the capability of flow cytometry in processing multi-parameter, high-dimensional data while ensuring timeliness and accuracy, this study proposes an intelligent flow cytometry analysis system. The system encompasses hardware design, software architecture, and algorithmic frameworks for flow cytometry. At the hardware level, a real-time data acquisition system was developed based on the cooperative operation of FPGA and ARM. On the software side, an embedded system with a Linux-based architecture was constructed, incorporating a suite of preprocessing, parsing, and batch normalization methods. For intelligent flow cytometry data analysis, a self-organizing mapping algorithm was introduced for dimensionality reduction, combined with an improved residual network from the field of deep learning, resulting in the development of an SE-ResNet-50 deep convolutional neural network model. Experimental results demonstrate that the SE-ResNet-50 model achieves a 4% improvement in overall accuracy and a 3.8% increase in precision compared to the original ResNet-50. The collaborative workflow integrating SOM and SE-ResNet-50 effectively processes the vast amounts of high-dimensional data acquired by flow cytometry. The findings validate the superiority of the proposed approach.

Abstract:With the rapid development of rail transit, the detection of track defects has become crucial for ensuring safety. This paper systematically reviews common types of track defects, such as fatigue cracks, burns of rails, and fastener looseness. It elaborates in detail on detection technologies including ultrasonic, eddy current, magnetic flux leakage, and machine vision, as well as their principles, applications, and advancements. This encompasses various derivative methods of ultrasonic detection, such as conventional ultrasound, phased array ultrasound, laser ultrasound, and ultrasonic guided waves. Additionally, it covers the innovations in eddy current detection regarding the suppression of the skin effect and combination with thermal imaging; the improvements in magnetic flux leakage detection in terms of signal processing and new lift-off layer design; and the characteristics of traditional image processing and deep learning methods in machine vision detection. Meanwhile, the application achievements of multi-source information fusion technology in track defect detection are expounded. For example, defect identification and localization are realized by collecting data from multiple technologies and integrating deep learning models. Finally, the challenges faced by multi-source technology fusion are analyzed, and suggestions for future research directions are proposed, providing a comprehensive reference for the development of track defect detection technologies.

Abstract:To address the issues of unstable motion of the robotic manipulator and frequent collisions with line fittings caused by frequent start-stop actions of the joints during obstacle negotiation,this paper proposes a collision-free line grasping trajectory planning method based on blended seventh-order non-uniform B-spline curves. First, an obstacle negotiation model for the robot is established, describing the mapping relationship between the pose of the walking wheel assembly and the joint space coordinates. The collision-free regions in the joint space are determined based on the principles of obstacle negotiation, and selection rules for collisionfree intermediate path points are developed. Second, seventh-order non-uniform B-spline curves are employed to fit the intermediate path points, constructing a collision-free trajectory with high-order continuity and controllable boundaries. Finally, co-simulation using Adams and Matlab, combined with the Non-dominated Sorting Genetic Algorithm(NSGA-II) algorithm, is performed to optimize the manipulator′s end-effector time and impact for the obstacle negotiation trajectory. The results demonstrate that the proposed method effectively avoids collisions with line fittings during obstacle negotiation and line grasping. Additionally, the acceleration and jerk variation curves of all joints are smooth without sharp peaks or abrupt changes. Compared to the linear-polynomial trajectory interpolation method, the proposed approach reduces the acceleration, jerk, and average jerk of the telescopic joint by 18.1%, 83.01%, and 78.32%, respectively, resulting in smoother robot motion and enhanced safety during obstacle negotiation.

Abstract:This paper addresses the issue of oil-water two-phase flow measurement under complex downhole conditions by developing an ultrasonic Doppler flow meter with sound velocity compensation. By establishing a one-dimensional velocity profile measurement model based on the Doppler effect and a sound velocity measurement model based on pulse echo intensity, and combining these with the principle of layered integration, a sound velocity compensated flow measurement model is constructed to achieve adaptive reconstruction of the velocity profile and high-precision flow measurement. On this basis, ultrasonic transducers and high-speed excitation and reception control boards suitable for high-temperature and high-pressure downhole environments are designed, and digital signal processing technology is used to achieve online demodulation of flow. Additionally, to ensure reliable operation under downhole high-temperature and high-pressure conditions, the measuring pipe section has been structurally and sealingly designed. A downhole flow test prototype has ultimately been developed. Experimental results indicate that the flow measurement module has a measurement error of less than 1% in the laboratory environment, can rapidly respond to fluid fluctuations, and can operate stably under extreme downhole conditions of 125℃ and 60 MPa. This technology can be widely applied to downhole measurement and adjustment scenarios and can be integrated into intelligent measurement and control systems, providing technical support for the construction of smart oil fields.

Abstract:Due to the volatility and randomness of photovoltaic power generation, it is difficult for traditional models to accurately predict it. To solve this problem, a prediction model of AW-CNN-LSTM is established based on clustering. First, the photovoltaic power plant historical data set is pre-processed and clustered using the K-means clustering algorithm based on the elbow method; secondly, an adaptive weight is established based on the distance between the training samples and the feature center of test samples of the same clustering category; then, an AW-CNN-LSTM network model suitable for different clustering categories is established based on the clustering results and adaptive weights. CNN are used to capture the relationships between different features, while LSTM are used to capture temporal features. Finally, the forecast results of each model are integrated to get the final forecast results. Experiments on the data set of photovoltaic power stations in the Australian Desert Solar Energy Research Center demonstrate the effectiveness of the proposed method.

Abstract:Aiming at the problem that modern industrial systems tend to focus on their predictive performance while paying little attention to equipment maintenance decision-making, a data-driven dynamic predictive maintenance method is proposed to avoid sudden system failures and ensure safe operation. First, the health status of the turbofan engine is monitored in real-time to obtain operating data, which is used to establish a turbofan engine remaining useful life model based on a convolutional neural networks-bidirectional gated recurrent unit-attention mechanism. The hyperparameters of the CNN-BiGRU-A are optimized using the black hawk optimization algorithm; second, the monitored data is input into the trained integrated network, and a dynamic predictive maintenance strategy with uncertain system task cycle is proposed based on the predicted remaining useful life; finally, the proposed method is verified by using the C-MAPSS data set to show that it can improve equipment predictive performance and perform good predictive maintenance afterward.

Abstract:Aiming at the UAV aerial image detection task, there are problems of tiny target size and complex background environment, which often lead to leakage and misdetection, this paper proposes a small target detection algorithm WT-YOLO based on YOLOv11 for aerial images. First of all, taking into account the problem that UAV aerial images are generally small targets, the structure of the YOLOv11 necking network is adjusted, and the output feature map is changed size, which improves the algorithm's ability to detect small targets. Secondly, the structure of Bottleneck and C3k2 module, named C3k2-WT, is redesigned in combination with WTConv to realize the efficient extraction of features. Again, Focal-Modulation is introduced to replace SPPF, which makes the model more robust in dealing with complex scenes by focusing and modulating the features at different spatial scales; finally, the shared convolution detection head is designed to reduce the number of parameters of the model through the convolution sharing mechanism, while enhancing the global information fusion capability between feature maps. The experiments of the improved algorithm on the VisDrone2019 dataset show that compared with the base YOLOv11s model, the accuracy (P), recall (R), and detection precision (mAP50) are improved by 5.6%, 5.9%, and 7.5%, respectively, and the number of params decreases by about one-fourth, which shows a good performance compared with other algorithms.

Abstract:In response to the issues of background interference in fine-grained images and the challenge of identifying the most discriminative features in the target region, this paper proposes an improved CVT-based fine-grained image recognition algorithm. First, a target region localization module is introduced into the CVT model. This module extracts features of the target region using a multilevel feature aggregation method and determines the target region via threshold-based decision-making. The original image is then cropped proportionally to reduce the interference of background information. Furthermore, a mechanism called MDCSAIA (Multi-Dimensional Channel Spatial-Aware Interaction) is proposed. This mechanism employs dimensional transformation to facilitate effective interaction between spatial information of adjacent channels and channel information of adjacent spatial positions, thereby enhancing the network′s ability to perceive the local details of the target region. Experimental results show that, compared to baseline algorithms, the proposed method improves recognition accuracy by 2.1%, 1.7%, and 1.5% on the CUB-200-2011, Stanford Cars, and Stanford Dogs datasets, respectively. These results validate the effectiveness of the proposed approach.

Abstract:In the Green Flexible Job Shop Problem (GFJSP), the complexity of production processes leads to low efficiency. Effective scheduling of Automated Guided Vehicles (AGVs) for transportation can ensure both production efficiency and cost control. This paper proposes a Multi-step Deep Double Q-Network (D4QN) algorithm to address the scheduling of green workshops and AGVs. The method first designs a mathematical framework based on the Markov Decision Process (MDP) to enable interaction between AGVs and the workshop. By adjusting the decision-making in real time through state features, action space, and reward functions, the algorithm coordinates job and AGV scheduling. Next, the algorithm for training decision-making is optimized, improving the calculation of Q-values and deep network training to obtain suitable solutions. Finally, two validation experiments are conducted to evaluate the learning performance of the proposed algorithm. The first experiment involves single-objective flexible job shop scheduling with the objective of minimizing makespan. The algorithm is trained on the Brandimarte and Kacem benchmark problems, and the results are compared with those of other deep learning algorithms. The results show that the proposed algorithm reduced the average processing time by 5.1~17.2 s and decreased the average optimal gap ratio by 7.5%~21%, demonstrating the algorithm′s superiority and stability. The second experiment focuses on multi-objective AGV scheduling in a workshop, with the goals of minimizing makespan and AGV energy consumption. The optimal number of AGVs is calculated for the MK01 problem instance, with results showing that four AGVs achieved a 3%~31.8% improvement in the normalized index compared to other quantities, proving its effectiveness in reducing costs and improving efficiency in the workshop.

Abstract:Currently, defect detection of CNC machine tool bearing seats primarily relies on manual visual inspection, which cannot meet the demands for high precision, high efficiency, and low error rates in industrial production. To address these issues, a defect detection algorithm based on an improved YOLOv5s is proposed for CNC machine tool bearing seats. Firstly, the HardSwish activation function is used to replace the GELU in ConvNeXtv2, and a novel CSCConvNeXtv2-HS structure is introduced, incorporating the CSC module to replace the C3 module in the backbone network. This modification reduces computational complexity while enhancing the feature extraction capability of key information. In the Neck network, a Scale Sequence Feature Fusion module is introduced to improve the model′s ability to extract multi-channel information. Finally, The final proposal of the Focal-Inner Loss loss function not only improves the speed of training convergence but also reduces the impact brought about by the imbalance in class distribution. Experimental results show that the improved model achieves an accuracy of 91.09%, a recall rate of 81.97%, and a mean Average Precision of 84.40%, with a processing speed of 61.73 fps. All evaluation metrics show improvements of 2.52%, 4.47%, 6.7%, and 1.12 fps compared to the original YOLOv5s model, thereby satisfying industrial production requirements.

Abstract:Given the difficulty and low accuracy of small target detection in aerial images, YOLOv8s improved small target detection method, namely YOLOv8-ERD, was proposed. Initially, the YOLOv8s Neck network was optimized using the Efficient Neck feature fusion method to facilitate the efficient merging of high-level semantic information with low-level spatial information. Subsequently, the receptive-field attention convolutional operation RFAConv was incorporated to emphasize the significance of various features within the receptive field slider and to bolster the feature extraction capability. Additionally, the C2f module was replaced with the improved DyC2f module, which employs dynamic convolution DynamicConv, thereby not only reducing computational overhead but also enhancing model performance. Lastly, a small target detection layer was added to refine the recognition accuracy of diminutive targets. Experimental results show that on Visdrone2019 public data set, compared with the benchmark model YOLOv8s, the mAP@0.5 of YOLOv8-ERD has increased by 5.0%, and the accuracy P has increased by 4.0%, and it performs well in comparison with other mainstream target detection methods.

Abstract:To address the existing PCB defect detection methods with high miss detection rate,poor generalisation and difficulty in balancing detection accuracy and speed, this paper proposes a PCB defect detection algorithm YOLOv8-CSM based on the improved YOLOv8n model. Firstly, a CoordAttention module is added at the end of the backbone network, which suppresses the influence of the complex background on the defective region of the PCB in order to improve the model′s detection accuracy; second, three SEAM modules are referenced in the detection header to expand the model receptive field and improve the model′s ability to identify small defects to reduce the miss detection rate; finally, MPDIoU is used to replace the traditional CIoU loss to optimise the regression effect of the bounding box and improve the convergence speed of the model. The experimental data show that YOLOv8-CSM can better balance detection accuracy and speed, and it is more generalizable. Compared with the base model, the Recall, Precision, mAP50 and FPS are improved by 4.3%, 1.8%, 2.7% and 42.76, respectively, which significantly enhances the model′s performance in PCB defect detection tasks.

Abstract:In the quality control of cigarette production, achieving precise detection of four types of tobacco shred (tobacco silk, cut stem, expanded tobacco silk, reconstituted tobacco shred) blending ratios has emerged as a critical technical challenge. To address the detection difficulties arising from subtle morphological variations and prevalent overlapped distributions of tobacco shreds, this study proposes a rapid overlapped tobacco shred segmentation algorithm based on an enhanced YOLOv8 framework. The method reconstructs the backbone network using a Res2Net architecture to amplify feature extraction capabilities for minute and complex patterns, while integrating ContextGuidedBlock (CGB) modules into the neck network to enhance boundary recognition accuracy in overlapped regions. Experimental results demonstrate that the improved model achieves notable performance metrics of mAP50 (86.5%), mAP50-95 (67.8%), and recall rate (81.9%) while maintaining real-time processing speed at 67 fps. Through ablation studies and comparative analyses with mainstream segmentation networks, the effectiveness and performance advantages of the proposed architectural modifications are rigorously validated. This algorithm not only improves segmentation precision but also optimizes frame rates, demonstrating superior applicability in practical production line environments.

Abstract:Blood cell detection is a crucial task in clinical diagnosis. However, due to the diverse cell types, significant size variations, frequent target overlap, and complex backgrounds, existing detection models still face challenges in terms of accuracy and robustness. To address these issues, this paper proposes an improved YOLOv8-based object detection model, YOLO-BioFusion.The model incorporates the ACFN module to enhance the detection of small and overlapping targets. Additionally, the C2f-DPE and SPPF-LSK modules are introduced to strengthen multi-scale feature fusion and extraction, improving the model′s robustness and generalization ability. Meanwhile, the adoption of the Inner-CIoU loss function accelerates model convergence and enhances localization accuracy.Experimental results on the BCCD dataset demonstrate that YOLO-BioFusion achieves an mAP@0.5 of 94.0% and an mAP@0.5:0.95 of 65.2%, outperforming YOLOv8-n by 1.9% and 32%, respectively. Moreover, with a computational cost of only 6.8 GFLOPs, YOLO-BioFusion exhibits great potential for applications in resource-constrained environments. This study provides an efficient and accurate solution for blood cell detection in complex backgrounds.

Abstract:Precision measurement technology has always been the core technology in the fields of biomedical impedance analysis, electrochemistry, power transmission, radio frequency antenna analysis and many other analytical instruments.In modern precision measuring instruments, the number of resistance-capacitance is often hundreds of thousands, and the impedance parameters are very strict, the quality of the resistance-capacitance used and the stability of its own impedance value is an important factor affecting the performance of electronic systems, so the precision resistance-capacitance measurement has become an important part of the design of many electronic systems.Under this background, a high precision resistance-capacitance analyzer based on hardware orthogonal phase-locked structure is proposed in this paper.The design takes hardware orthogonal phase discrimination as the core. Around this part, the orthogonal frequency generator, I-V conversion regulation module, Sallen-Key filter and other hardware structures are designed. At the same time, the MCU system is designed for module control and drive, task allocation and scheduling, and data acquisition and processing. Finally, performance tests were conducted on the key components of the system and the physical complete machine, achieving the judgment of resistance and capacitance as well as the measurement of impedance values. After the actual system test and comparison with the standard bridge test, the average value of the resistance measurement error of this system is 0.032 4%, and the average value of the capacitance measurement error is 0.054 7%. And while achieving a lower measurement error, it ensures the small volume and portability of the overall system, ultimately enabling the design to reach the expected goal.