Abstract:In the field of oil and gas development, the sealing performance test of oil casing after installation is particularly important.Torque sequence data is an important basis for judging the sealing performance of the oil casing, which can be used to judge whether the buckle is qualified. In order to identify and classify the sealing performance of the oil casing by using the information of the buckled torque sequence data, a new network model was built which named PSE-TCN network based on the TCN model integrated with position encoding and self-attention mechanisms. By comparing the accuracy of results under different strategies, the learning process of the model was demonstrated. The effectiveness of this method was validated by comparing it with other network models. Experimental results show that torque sequence recognition accuracy was significantly improved by the PSE-TCN network compared with other classical network models and several improved TCN models. The recognition accuracy of this model achieved 93.41% on the self-made UCR_whorl dataset.

Abstract:To enhance the stability of angle measurement methods based on single-vision techniques, which are susceptible to random disturbances from environmental or systemic sources, we propose a dual-viewpoint visual angle measurement method based on complex-valued neural networks. Feature extraction is conducted manually, followed by an assessment of the features’ relevance and monotonicity with respect to angles to facilitate feature selection. To address the significant numerical discrepancy between the 0° and 360° labels, which impacts training outcomes, angles are represented using Euler′s formula. This representation facilitates the construction of a complex-valued neural network with both complex inputs and outputs for angle computation. Experimental results demonstrate a significant improvement in measurement accuracy; the proposed method reduces the mean error by 0.322° and the root mean square error by 0.64° compared to methods based on deep neural networks using a single viewpoint, maintaining high performance across various environmental test sets. By leveraging the robustness against environmental disturbances provided by dual viewpoints and the strong fitting capabilities of complex-valued neural networks for angle labels, this model enhances the accuracy and stability of radial visual angle measurements while adhering to the constraints and stability of mathematical models.

Abstract:Aiming at the problem of easy interference of electrostatic signal and low fault recognition rate when the new electrostatic monitoring technology is applied to rolling bearing fault diagnosis, a method of electrostatic signal recognition of rolling bearing fault based on the combination of Bayesian optimization SVM is proposed. First of all, through the electrostatic simulation test platform constructed, the electrostatic signals of different wear states high speed are collected, and the feature sets of different working conditions are selected according to the timedomain feature parameters; and then the hyper-parameters of the minimum error of SVM are selected using Bayesian optimization to achieve the effect of completing the diagnostic model training, and the diagnostic accuracy of the models is evaluated with the results of the confusion matrix after training. The research results show that this method has certain recognition ability for bearings with different fault characteristics under electrostatic monitoring, and the Bayesian optimization algorithm can effectively improve the recognition efficiency, and its average recognition accuracy can reach 98.82%.

Abstract:In the classification and recognition of motor imagery EEG features for limb movements, there exists a problem of low action recognition accuracy when fusing features from different domains. To address this issue, this study designs an EEG-symmetric positive definite network model for motor feature classification, tailored to the complex cross-domain relationships of motor imagery EEG features in multi-channel data collection. This model effectively extracts and integrates features from different domains, achieving accurate classification of limb features and action recognition based on EEG signals. Experimental results demonstrate that on the BCI Competition IV 2a dataset, which contains motor imagery data of four types of limb movements, the proposed classification model achieves an action recognition accuracy of 0.85 and a Kappa coefficient of 0.80, indicating high precision.

Abstract:The application formula and process of tobacco fragrance are the core technology of the tobacco industry. In China, each tobacco industry has chosen the construction of fragrance categories as the next round of strategic choices. Its differentiation is a technical key point for the competition among various cigarette brands. This paper proposes a machine vision method combining dual light source illumination to solve the problem of poor quality judgment by manual judgement in the processing of tobacco fragrance configuration and preparation, and designs and manufactures an appearance quality qualification detection device for tobacco fragrance based on this. Using white light and red light as the main test light sources and green light as auxiliary detection light source, a dual light source coaxial forward illumination environment is set up; by fixing the optical plate for lighting and image acquisition module as a whole and combining the slide table with the stepping motor to rotate and stop at designated points, the machine vision method is used to eliminate reflections and automatically analyze color model parameters and detect the appearance quality qualification of the tobacco fragrance. The results show that the relative standard deviation of the parallel test of single tube sample image is less than 0.9968%, and the relative standard deviation of the parallel test of the same batch sample is less than 0.021 7%. The experimental results show that the precision and repeatability of the instrument are good, and can provide support for further promoting the intelligent management of the tobacco fragrance configuration detection industry.

Abstract:In intelligent networks, multiple service flows have different transmission requirements in terms of delay and bandwidth, and the burstiness of self-similar traffic exacerbates delay and packet loss rate. To address this problem, an improved WFQ scheduling algorithm based on traffic prediction (LPR-WFQ) is proposed. This algorithm uses the TLGP strategy to classify traffic based on the mean and variance of traffic. Based on the Bayesian estimation idea, it predicts future traffic levels by calculating conditional transition probabilities. The weights are dynamically adjusted based on the prediction results and the mean arrival rate, thereby reducing delay and packet loss, improving service quality, and optimizing the calculation method of virtual finish time. Simulation results show that compared with other scheduling algorithms, this algorithm improves the delay, delay jitter, throughput and packet loss by 6.01%, 9.66%, 5.37% and 38.57% respectively, indicating that the algorithm can meet the performance requirements of differentiated services.

Abstract:Aiming at the multi-point path planning problem of mobile robots, a path planning algorithm combining ant colony algorithm and bat algorithm is proposed in this paper. The ant colony algorithm is used to establish the shortest path network between nodes. The pointing angle and turning angle are introduced as heuristic information in the traditional ant colony algorithm to reduce the paths′ turning times and turning angles. The reward and punishment mechanism is used to optimize the pheromone updating mode and improve the convergence speed of the algorithm. The objective function of multi-point path planning is based on the shortest path network. When solving the optimal node access order, the structure of the bat algorithm is improved, the hierarchical search method and a new local optimization mechanism are introduced, and the bat algorithm′s solving accuracy, speed, and stability are improved. The simulation results demonstrate that the proposed algorithm effectively addresses the issue of multi-point path planning. In comparison to existing algorithms, it exhibits lower computational complexity, higher search efficiency, smoother overall paths, and shorter lengths.

Abstract:To address problem that the ORB-SLAM2 algorithm is prone to mismatching and cannot build a dense map during feature matching, the GMS algorithm is introduced to improve the mismatching problem in the ORB-SLAM2 algorithm and add a dense map thread. First, an image pyramid is established, and a grid division is performed on each layer of the image pyramid to extract feature points. A four-tree strategy is introduced for feature point selection in each grid, resulting in a uniform distribution of feature points. Second, the GMS algorithm is introduced in the feature matching stage to eliminate false matches. Finally, the dense point cloud map is built based on the pose estimation and key frames. Through the experimental verification on TUM data set, the results show that the matching number of the improved algorithm is 7.82% higher than that of the original ORB-SLAM2 algorithm, and the matching time is reduced by 8.53%. The improved algorithm is applied to the automatic navigation and obstacle avoidance of mobile robot, which can improve the reliability and operation efficiency of the system.

Abstract:It is essential to accurately predict the state of health (SOH) of lithium-ion batteries. Aiming at challenges such as differences in degradation mechanisms at different stages of a single battery cycle and incomplete data acquisition in practical utilization scenarios, a lithium-ion battery SOH estimation method based on Involution-Vision Transformer (IViT) is proposed. Features that can effectively characterize the degradation information of lithium-ion batteries are automatically extracted from the voltage-time profile, weights are adaptively assigned at different positions using the Involution module, and Vision Transformer is used to learn the high-level feature representations at different stages and capture the global dependencies. The experimental results show that the prediction error of IVIT is around 0.5%, and the error is only around 2% when the overall data is missing 50%, proving the effectiveness and stability of the proposed method.

Abstract:In the surface measurement of metal workpieces based on line structured light, this paper proposes a laser stripe centerline extraction algorithm based on improved principal component analysis to address issues such as strong reflection and laser stripe breakage on the surface of metal workpieces. Firstly, for the irregular reflection of metal workpiece surface, the optical fringe region of image was extracted based on maximal variance between clusters (OTSU); Secondly, in response to the problems of high convolution operations, low efficiency, and poor real-time performance of the Steger algorithm, an improved Steger algorithm based on principal component analysis (PCA) was proposed. The covariance matrix of the gradient vector was constructed using PCA to estimate the normal direction of the stripe, and the second-order Taylor expansion was used in this direction to obtain accurate sub-pixel coordinates of the stripe center. The experimental results show that the algorithm proposed in this paper can effectively extract laser stripe areas under severe reflection conditions on the surface of metal workpieces. At the same time, the standard deviation of the extracted laser stripe centerline is reduced by about 0.25 pixels compared to the grayscale centroid method, and the speed is increased by nearly 13 times compared to the Steger algorithm. It can quickly and accurately extract the laser stripe centerline, meeting the realtime detection requirements of structured light 3D vision.

Abstract:Aiming at the problems of poor structural consistency and insufficient texture details in the inpainting results of existing image inpainting algorithms, an image inpainting algorithm based on multi-feature fusion was proposed under the framework of generative adversarial network (GAN). Firstly, the dual encoder-decoder structure is used to extract the texture and structure feature information, and the fast Fourier convolution residual block is introduced to effectively capture the global context features. Then, the information exchange between structure and texture features was completed through the attention feature fusion (AFF) module to improve the global consistency of the image. The dense connected feature aggregation (DCFA) module was used to extract rich semantic features at multiple scales to further improve the consistency and accuracy of the inpainted image, so as to present more detailed content. Experimental results show that, compared with the optimal comparison method, the proposed algorithm improves PSNR and SSIM by 1.18% and 0.70% respectively, and reduces FID by 3.99% on the Celeba-HQ dataset when the proportion of damaged regions is 40%~50%. On the Paris Street View dataset, PSNR and SSIM are increased by 1.17% and 0.50%, respectively, and FID is reduced by 2.29%. Experimentally, it is proved that the suggested algorithm can effectively repair large broken images, and the repaired images have a more sensible structure and richer texture details.

Abstract:In order to further improve the compatibility of machine vision systems and enrich the types of video formats processed by encoding and decoding systems, a multi interface video encoding and decoding system based on FPGA was designed. By using the asynchronous DDR read-write principle to build the codec selection module and complete the conversion operation of different video formats, the final system supports the decoding of PAL, HDMI and Cameralink videos as well as the encoding functions of HDMI, Cameralink and LVDS videos. Meanwhile, by comparing the transmission characteristics of different video interfaces, the seamless conversion between the above video interface standards is realized. The system can not only be used as an independent video codec system, but also can be connected to ARM processor through LVDS interface, thus expanding its application scenarios. Experimental results show that the system can accurately decode PAL video with a resolution of 720×576, Cameralink video with a resolution of 640×512 and HDMI video with a resolution of 1 080p, and then output it through HDMI, Cameralink and LVDS video interfaces respectively. In addition, the consumption of all kinds of resources in the system does not exceed 50%, which ensures the efficient operation of the system.

Abstract:To address the frequent disassembly required for program updates in storage systems within the industrial testing field, as well as the unique need for data storage without a host computer, a smart online upgrade storage system design based on FPGA is proposed. This system utilizes FPGA as the main controller and employs a combination of Gigabit Ethernet and FLASH. Update instructions and configuration files are transmitted via Gigabit Ethernet to the program memory, where they are partitioned, erased, and written to SPI Flash, enabling online upgrades of the FPGA program. Additionally, the optocoupler instruction parsing module enables the system to operate independently of a host computer, performing intelligent data storage autonomously. Furthermore, the system integrates a reliable feedback design using a custom DR_UDP protocol, optimizing the communication efficiency and stability of the Gigabit Ethernet port. Functional verification analysis confirms that the system operates stably, flexibly, and reliably, with a Gigabit Ethernet transmission rate reaching 700 Mb/s, and no data loss detected. This system can be widely applied in various scenarios where disassembly is inconvenient.

Abstract:Spatio-temporal graph convolutional network (STGCN) enhances the accuracy of traffic speed prediction by capturing the spatial dependencies and temporal dependencies in traffic data through graph convolution and time convolution. However, the hardware implementation of traffic speed prediction using STGCN faces challenges such as high computational demands that do not meet the real-time requirements of practical applications and high resource consumption leading to increased costs. To optimize the traffic speed prediction STGCN model, a method for optimizing the FPGA implementation structure combination of traffic speed prediction STGCN is proposed. Initially, the model is optimized through lightweight pruning and precise selection of prediction data bit-width to reduce computational complexity and resource consumption, verified by Python simulation for feasibility. Subsequently, an optimization strategy using pipeline, parallel computing, and alternating data stream storage is introduced to enhance system computational speed. Finally, the traffic speed prediction STGCN is implemented and tested on FPGA using Verilog programming. Experiments with the PeMSD7(M) dataset show that the FPGA implementation reduces the time for single data traffic speed prediction to 355.5 μs, maximum processing speed increases of 25.9×, 6.7× and 3.5× compared to CPU, GPU platform and FPGA design option 1 comparisons, respectively, proving that the proposed method significantly improves processing speed while maintaining prediction accuracy.

Abstract:Deep learning has become one of the important tools for hyperspectral image classification due to its modular design and powerful feature extraction capability. However, effectively extracting deeper features and simultaneously improving the analysis of spatial and spectral joint features remains an urgent challenge. In response to these issues, a deep feature extraction residual network is proposed in this paper, composed of two key components: a multi-level transfer fusion residual network and a spatial-spectral multi-resolution fusion attention residual network. The multi-level transfer fusion residual network effectively promotes interaction between feature information to obtain deeper-level features. Subsequently, the spatial-spectral multi-resolution fusion attention residual network ensures comprehensive extraction of spatial-spectral joint features and multi-resolution features from hyperspectral data. To validate its effectiveness, the performance of the proposed method was evaluated on three hyperspectral datasets, Indian Pines, Pavia University, and Salinas Valley, achieving classification accuracies of 98.10%, 99.81%, and 99.94% respectively. Experimental results demonstrate that, compared to other methods, this network exhibits better generalization capability and classification performance.

Abstract:In order to realize unmanned production in factories, textiles need to be sorted efficiently. The manual classification method for traditional textile production plants has the problem of low efficiency and difficulty in meeting the needs of large-scale production. Artificial intelligence and computer vision advanced technology were applied to textile material classification, and a textile material classification algorithm based on DSCI-YOLOv8 was proposed. On the basis of the original classification network of the YOLOv8 model, the coordinate information attention module is added to enhance the model′s ability to extract the features of textile materials at different scales, improve the accuracy of network classification, and reduce some of the calculations and parameters required for calculation. Secondly, the distributed offset convolution is added to the C2f network module, which improves the network structure of the classification neural part, so that the memory usage is reduced and the computation speed is improved. Experimental results show that the accuracy of the improved model is increased by 2.09 percentage points and 13.5% increase in image processing per second compared with the YOLOv8 model. While greatly reducing the calculation cost, it effectively improves the accuracy and speed of textile material classification. It can meet the testing needs of the textile industry for product category classification and quality.

Abstract:To address the problems of low accuracy, easy false detection and missed detection in the existing insulator self-explosion defect detection methods under complex backgrounds and foggy environments, an improved YOLOv8 insulator self-explosion defect detection algorithm is proposed. First, the SPD-Conv module for low resolution image and small target detection is introduced into the backbone network to fully extract the feature information of insulator defect target. Secondly, BiFPN is integrated with the SimAM attention mechanism to build the BiFPN_SimAM module, replacing the concat connection of PANet to achieve multi-scale feature fusion and enhance the overall performance of the network. The experimental results show that the precision and mAP@0.5 of the improved algorithm for insulator self-explosion defect detection reach 95% and 93.1%, respectively, which are increased by 1.8% and 1.5% compared with the original YOLOv8 algorithm, and it also has a good detection effect on insulator self-explosion defect detection under complex background and foggy environment.

Abstract:To address the high proportion of small target objects in waterborne debris detection, the interference caused by multiple factors such as water surface fluctuations and shoreline reflections, and the high demands on device performance due to the large number of parameters and computational load of detection models, we propose a lightweight, high-precision, real-time detection model, LS-YOLO. First, this algorithm uses the HS-FPN pyramid network design to construct the Neck network structure of YOLOv8. The constructed network structure sacrifices a small part of the accuracy and significantly reduces the number of parameters and computational complexity of the model. Secondly, HS-FPN is improved by introducing the CAA context-anchored attention mechanism to capture remote contextual information to improve detection accuracy. Then, by replacing the loss function with Wise-IoUv3, which features a dynamic focusing mechanism, the detection performance is significantly improved, increasing the robustness of the model. Finally, LAMP pruning technology is used to prune the model to reduce the number of parameters and calculations of the model. The experiment shows that the improved LS-YOLO has a 0.9% increase in mAP50 compared to the baseline model, a 3.2% increase in recall, a reduction in parameters to 19.83% of the baseline model, a reduction in computational cost to 44.44%, and a reduction in model size to 22.22%. The optimized detection algorithm not only significantly improves detection performance and feature extraction accuracy, but also facilitates deployment on resource-constrained hardware platforms.

Abstract:Addressing the issue of subpar performance in identifying shallow water marine life in underwater environments using existing methods, we propose an improved method based on the RT-DETR benchmark model. Initially, the reparameterization network RepViT is utilized as the backbone of the model, enhancing its feature extraction capabilities. Subsequently, a reparameterized parallel dilated convolution (RepPDC) is constructed and incorporated into the neck network, enabling the model to effectively capture long-range contextual information, thereby improving the model′s recognition accuracy. Lastly, a bidirectional feature fusion module (CAFM) is constructed based on the attention mechanism, enhancing the model′s ability to focus on key information in underwater environments. Experimental results demonstrate that the improved method significantly boosts the mAP50 to 87.5%, mAP75 to 70.9%, and mAP50:95 to 64.9%, with fewer parameters, making it a promising candidate for practical applications in the identification of shallow water marine life.

Abstract:To address the challenges in food image recognition caused by small inter-class differences, large intra-class variations, and complex structures, this paper proposes a food image recognition method that integrates multi-scale features and an attention mechanism. First, the ConvNeXt model, which has stronger feature extraction capabilities, is used as the backbone network to better capture the detailed features of food images. Next, an improved ASPP module is introduced to expand the receptive field and utilize multi-scale information, enhancing the model′s ability to capture features at different scales. Finally, an attention mechanism is added after each convolutional block to improve feature representation and the ability to capture contextual information. Experimental results show that the proposed method achieves accuracies of 91.56% and 87.22% on the extended Vireo Food172 dataset and the ETH Food101 dataset, respectively, which represents an improvement of 2.05% and 1.66% over the original model, thus verifying the effectiveness of the proposed method.

Abstract:To address the challenges of small defect targets, multi-scale issues, and high reflectivity on the surface of the inner car door handle, we first tackle the problem of defect features being obscured during image acquisition due to surface curvature and mirror reflection by using a bowl-shaped light source and reducing the angle of the image acquisition surface. Then, recognizing the limitations of traditional RT-DETR models, such as poor detection accuracy and slow speed, we propose an improved RT-DETR object detection method. This method builds upon the RT-DETR framework, utilizing parallel dilated convolutions and the CA attention mechanism combined with convolutional re-parameterization in the backbone network to increase the receptive field and establish long-distance semantic information while improving the network inference speed. Additionally, extra detection layers are added to improve the network′s feature extraction capability for small object detection. In the multi-scale feature fusion stage, we use an improved BIFPN structure to enhance the model′s information interaction capability. Finally, ablation experiments show that, compared to traditional RT-DETR-based detection methods, our proposed improved RT-DETR method increases the mean Average Precision by 6.5%, achieves a detection speed 1.6 times that of the traditional model, and reduces the model′s parameter count to only 76.5% of the original network, validating the effectiveness of our proposed method.

Abstract:The railway signal system is an important technical means to ensure the safe and efficient operation of railway transportation. As a key equipment of the railway signal system, the completeness testing of the system itself is essential for the computer interlocking system. The interlocking human-machine interface is an important component of the interlocking system. Through the operation of the operators, control commands can be sent to the signal equipment, and on-site equipment status information can be received and displayed. Testing the interlocking human-machine interface according to standard specifications is an important technical means to ensure the normal operation of the interlocking system and ensure the safety of railway operations. At present, the testing of interlocking human-machine interfaces mostly relies on manual labor, which has problems such as low testing efficiency and untraceable testing processes. This article proposes a template matching scheme suitable for real-time graphical interface detection based on the normalized squared difference algorithm; analyze the local features of the interlocking human-machine interface image and propose a non-invasive and distortion free image pixel feature recognition method; modeling and abstracting manual operation steps into computer recognizable language; propose 13 custom keywords to simulate interlocking human-machine interface operations; automatically capturing and analyzing image, text, and speech information, accurately calculating the RGB primary color model values of the image, determining the compliance of test results with specifications, and improving the accuracy and consistency of detection results. After verification and comparison, the proposed interlocking human-machine interface detection method has achieved full process automation testing, visualized all operation processes, and traceable test results and intermediate links, greatly improving testing efficiency and the credibility of test results.