Abstract:Accurate detection of defects in wafer images is of great significance for timely identification of abnormal faults in wafer production. In the wafer testing phase, the deep learning method has been widely used in wafer defect detection due to its excellent feature extraction capability. However, traditional deep learning models often rely on a large number of adequately labeled and high-quality data for training, and in practical applications, balanced and sufficient labeled data is often difficult to obtain. To address this issue, we propose a VGGNet deep learning model that integrates an improved multi-head attention mechanism with a residual structure, aiming to extract more comprehensive features from imbalanced data sets to achieve accurate detection of wafer surface defects. Specifically, we use an improved multi-head attention mechanism to map the input wafer image features to multi-dimensional subspaces, which significantly improves the expressiveness and generalization performance of the model. At the same time, the residual connection is introduced into the full connection layer of traditional VGGNet, which effectively alleviates the problem of gradient disappearance in deep network training. To validate the effectiveness of the VGGNet with the improved multi-head attention mechanism and residual structure(RS), extensive experiments were conducted on the WM811K dataset, achieving a classification accuracy of 94.3%, which is 3% higher than the traditional VGGNet and 1% higher than existing similar models on average. The experimental results show that on the real data set WM811K, the proposed method not only improves the robustness of wafer defect detection, but also significantly outperforms the existing algorithms on the non-equilibrium data set.

Abstract:In semiconductor manufacturing, wafer map defect detection is crucial for the rapid localization and identification of defects, which is significant for enhancing wafer product quality and production efficiency. However, existing methods have limitations, such as overly complex models and excessively deep network structures that struggle to leverage multi-level features for accurate classification. To address these issues, this paper combines a Stem-Dense feature extraction module with a multi-scale attention feature fusion module to propose a novel network architecture—multi-scale defect detection network with enhanced feature extraction (MSD-DFE). MSD-DFE effectively captures rich shallow feature information through the dense connection structure of Stem-Dense and multi-scale attention-based feature fusion technology, while significantly reducing the number of parameters and computational complexity of the model. The multi-scale feature extraction module integrates wafer map information from various scales, enhancing the model’s ability to extract defect features. Additionally, the introduced attention mechanism allows the model to focus more on defect areas, thereby improving classification accuracy. Experimental results show that MSD-DFE achieves an average accuracy of 97.4% on the WM-811K dataset, outperforming current mainstream methods, indicating its high potential for practical application in industrial settings.

Abstract:The thermal barrier coating of aviation turbine blades can reduce the surface temperature of the blades and prevent high temperature corrosion on the surface. The coating defects affect the performance of the blades seriously. The complex blade surface shape and its matrix structure lead to the difficulty of non-destructive testing of coating defects. In view of the sensitivity of Rayleigh wave to damage changes such as surface stress and surface micro-cracks, a non-destructive testing method for micro-crack defects on the surface of turbine blade thermal barrier coating based on generalized Rayleigh wave propagation is proposed. A special ultrasonic array transducer was designed and a Rayleigh wave detection system was built. The Rayleigh wave detection signal propagating in the back coating along the width direction of the blade was extracted, and the correlation between the distribution characteristics of the blade amplitude and the complex structure inside the blade was analyzed. Finally, the influence of artificial narrow slots in the blade on the propagation characteristics of Rayleigh waves is analyzed, and a generalized Rayleigh wave detection method for coating defects is proposed. The results show that the complex surface profile and the inner cavity diversion structure of the turbine blade have a significant effect on the amplitude of the Rayleigh wave. The Rayleigh wave amplitude distribution of the intact blade coating is compared with the Rayleigh wave amplitude distribution characteristics of the blade coating with a narrow groove (500 μm×80 μm×20 μm). The results show that the ultrasonic amplitude propagating on the path with a narrow groove increases significantly, with an average increase of 54.3 mV. This feature can be used for non-destructive testing of micro-crack defect damage on the surface of turbine blade thermal barrier coating.

Abstract:To address the challenges of detecting multi-scale and deformed road defects in complex road scenarios, an improved YOLOv8n model for road defect detection, named DMS-YOLO, is proposed. First, an adaptive context-aware feature pyramid network is designed to achieve global fusion and dynamic weighting of multi-scale features, significantly enhancing the model’s ability to perceive and express complex defects. Compared to existing mainstream feature pyramid networks, this approach demonstrates clear advantages in both accuracy and computational efficiency. Second, an adaptive multi-scale dynamic detection head is introduced, leveraging deformable convolution (DCNv3) to improve the model’s capability in capturing complex shape features, and a Collaborative Attention Mechanism is designed to integrate scale and task attention, enhancing the model’s understanding of multi-scale information. Finally, the CIoU loss function is improved using the Focaler-IoU idea to enhance the detection of small targets. Experimental results show that, with reduced computational cost, the DMS-YOLO model achieves a mAP@0.5 of 87.9% on the RDD2022 dataset, a 3% improvement over the baseline model. The model has 3.67×106 parameters, 8 GFLOPs of computational cost, and a model size of only 7.3 MB, demonstrating its lightweight nature and ease of deployment. Additionally, on the SVRDD dataset, DMS-YOLO improves on all performance metrics, further validating the model’s generalization and robustness. Compared to other mainstream models and state-of-the-art detection algorithms, DMS-YOLO shows superior overall performance, demonstrating its practical application value in road defect detection.

Abstract:For the three-dimensional defect detection task of SOP chip pins, existing point cloud deep learning methods struggle to effectively detect common pin defects. To address this issue, a DCPP image is defined and a corresponding DCPP dataset is created. A DCPP-PointNet defect detection algorithm is also proposed, specifically designed for DCPP images. This algorithm incorporates a LSEF network, which enhances the model’s rotational robustness and ensures good detection performance even with rotated point cloud data. Additionally, a new iRMSC-Net network is designed to replace the feature encoder in PointNet++, improving the model’s ability to learn local edge features of point clouds and enabling precise classification and location of common SOP chip pin defects. Focal loss function is employed to tackle the imbalance between positive and negative samples, allowing the model to focus more on hard-to-distinguish defect samples and thus improving detection accuracy. Experimental results on the self-built DCPP dataset show that the DCPP-PointNet network surpasses existing point cloud segmentation models such as PointNet, PointNet++, and DGCNN in terms of OA and mIoU. It achieved an OA of 98.9% and an mIoU of 93.7%. Ablation studies further confirm the effectiveness of the improvements in DCPP-PointNet, where the combined action of the LSFE network, iRMSC-Net feature encoder, and Focal loss function significantly enhances the model’s detection accuracy and robustness.

Abstract:Construction cranes are the core equipment of modern engineering, and their high-risk operation at height is prone to cause major accidents and economic losses, seriously threatening safety. In order to improve the efficiency and accuracy of defect recognition and reduce the risk of operators climbing up to inspect, a surface defect intelligent detection method FRE based on UAV images is proposed. The surface defects of construction cranes are diverse, tiny in scale and complex in background, and the traditional YOLOv8 network is difficult to realize high-precision defect detection due to the lack of multi-scale feature fusion capability and the limitation of environmental adaptability. Utilizing the UAV inspection construction equipment, two typical lifting machine defect image datasets of wire rope defects and metal structure corrosion are established. The C2F module in the YOLOv8 backbone network is replaced with the RepViT Block module to improve the performance and efficiency of the model in image understanding and processing, which significantly reduces the computational complexity and latency, and the training speed is increased by 46.4% and 2.6%, respectively; the C2F module in the neck network is replaced by the FasterNet Block module, which improves the performance of the localization of defects and improves the ability of detecting small targets; the EMA module is embedded into the backbone network to suppress the interference of background information and make the model more focused on defect features. Compared with the existing defect detection, the detection accuracy of the model reaches 88.0% and 94.1%, respectively. Meanwhile, the number of model parameters decreased by 23.26% compared with the YOLOv8 model. The results show that the method can quickly and accurately detect the surface defects of construction cranes, which has certain social application value.

Abstract:In order to solve the problems of small defect area of industrial printed circuit boards (PCB), high false detection and missed detection rate caused by background interference, and difficult defect location, an improved circuit board defect detection algorithm based on YOLOv8n was proposed. First, by adjusting the feature fusion levels of the feature pyramid networks (FPN) in the backbone network, introduce a 160×160 tiny-target feature layer and detection head to replace the original 20×20 large-target feature layer and detection head., which enhances the network’s ability to extract features of small targets. Secondly, a parallelized patch-aware (PPA) attention module is introduced between the backbone and the neck. Through the multi-branch feature extraction part, it captures features of different scales and levels of the target, strengthening the model’s ability to extract and fuse local and global information. While avoiding the interference of complex background features, it also improves the utilization efficiency of the target feature information. Furthermore, the efficient multi-scale attention module (EMA) is introduced at the neck end to avoid more sequential processing and model depth, and at the same time, the cross-space learning ability of the network is enhanced. Finally, normalized wasserstein distance-efficient intersection over union) is employed as the bounding box regression loss function (NWD-EIoU). By introducing the normalized Wasserstein distance (NWD) to construct a geometrically-aware similarity metric, it alleviates the cumulative localization errors caused by minor offsets of detection boxes, improves the model’s positioning accuracy for micro-defects on PCBs, and accelerates convergence. The experimental results on the publicly available PCB defect dataset PKU-Market-PCB show that the mAP@0.5 of the improved method has increased by 4.2% compared with the original algorithm. The Precision and Recall metrics have increased by 7.7% and 4.3% respectively. Compared with the same type of single-stage object detection methods, the improved method meets the requirements of high-precision PCB defect detection.

Abstract:Urban rail transit track condition monitoring is one of the critical tasks for ensuring the safety of railway transportation systems. The urban rail transit track includes key components such as rails, fasteners, bolts, and sleepers. In response to the demand for real-time and refined detection, this study, building on previous work, further investigates and proposes an innovative intelligent method based on instance segmentation for the rapid and refined identification of multiple key components of urban rail transit tracks, analyzes, and quantifies the detection results of common defects. Specifically, this research, based on the existing RTLSeg model, integrates field-of-view enhancement and image post-processing techniques, proposing an improved track image segmentation and evaluation model (ABI-RTLSeg). Firstly, to enhance the model’s learning of high-level semantic information, this study introduces a dilated spatial pyramid pooling (ASPP) module into the deep backbone network. Secondly, a convolution-based bilinear interpolation structure is incorporated into the Coord-Protonet to obtain higher-quality prototype masks and semantic information awareness. Lastly, based on the visual features of defect segmentation masks, a segmentation result analysis module is constructed, employing ellipse fitting and morphological analysis methods to analyze the safety status of common defects. Experimental results demonstrate that this method is feasible for rapid and refined detection, segmentation, and analysis of multiple target key components and common defects of railway track lines, and its performance surpasses that of the comparative baseline models. In particular, ABI-RTLSeg is able to achieve 90.91% bbox mAP and 91.67% mask mAP with the customized dataset. Meanwhile, the average inference speed reaches 25.62 fps. The average detection accuracy and recall are 100% and 99.85%, respectively. Furthermore, the feasibility of the proposed methods for assessing the severity of fastener damage and estimating key parameters of rail corrugation has been explored through multiple case studies. In summary, this study provides a new technical approach for the intelligent monitoring of rail transit track lines, which is of great significance for improving the safety and reliability of the railway transportation system.

Abstract:Aiming at the issues of insufficient feature representation capability of latent vectors for positive samples, suboptimal reconstructed image quality by the decoder, and inadequate discriminative ability of the discriminator in the GANomaly model, a tire X-ray image defect detection method based on D2GANomaly is proposed. First, a multi-scale dynamic residual block (MDRB) is introduced into the encoder, which combines adjustable kernel convolution (AKConv) with residual connections to dynamically fuse multi-scale features and enhance fine-grained feature extraction capabilities. Second, a channel residual sub-pixel decoder (CRSD) is incorporated into the decoder section, utilizing dual decoders for parallel learning to optimize the reconstruction quality of complex textures and details. Finally, the discriminator employs a dual discriminative module network (DDMN), which uses switchable atrous convolution (SAC) to select the optimal dilation rate, thereby enhancing the model’s ability to detect defects of varying sizes in tire X-ray images and improving its discriminative performance. Experimental results demonstrate significant improvements in two core performance metrics, Area under the receiver operating characteristic curve (AUC) and average precision (AP). Compared to the original GANomaly model, the proposed method achieves a 13.7% increase in AUC and a 16.4% increase in AP. This indicates that the improved model effectively enhances the accuracy of tire defect detection.

Abstract:Aiming at the problems such as low echo signal-to-noise ratio, high sidelobes, and wide main lobes, an ultrasonic guided wave detection method based on Nonlinear Frequency Modulation （NLFM）-Barker coding excitation is proposed. Firstly, Barker code is chosen as the base encoding, which is then combined with sinusoidal signals, line frequency modulation (LFM) signals, and NLFM signals, respectively. Secondly, the time-domain, frequency-domain, and pulse compression characteristics of the three composite coded signals are analyzed through simulations. The simulation results show that after weighted matching filtering, the NLFM-Barker pulse compression signal has the narrowest main lobe width, which is 9.23 μs at -6 dB. Finally, to further verify the effectiveness of the proposed composite coding excitation, defects are simulated by attaching mass blocks of different sizes to a 2 000 mm long CHN60 rail. The experimental results show that when the rail is intact, the main lobe width of the NLFM-Barker signal is reduced by 5.9%, and the peak-sidelobe level (PSL) is decreased by 2.161 4 dB compared to the traditional Sin-Barker signal. When defects are present in the rail web and NLFM-Barker coded excitation is applied, the energy variation of the received ultrasonic guided wave signal becomes more distinct across different defect sizes. To sum up, this study provides a reliable and effective solution for rail web defect detection and quantitative analysis.

Abstract:In order to solve the problems of low detection accuracy and high complexity of existing models due to the variety of defect types, significant size differences, and high complexity of existing models in the detection of steel surface defects, a lightweight detection algorithm YOLOv8n-CSG with improved YOLOv8n was proposed. Firstly, the design of the CG Block module was introduced C2f_CG which enhanced the ability to capture the surrounding features and enhance the information relevance. Secondly, a C2f_Star module is designed by adding the Star Block module, which maps the input data to the high-dimensional nonlinear feature space and generates rich feature representations, which makes the model more effective in dealing with subtle defects. Finally, a lightweight detector GSE_Detect integrating GSConv and EMA attention mechanisms was designed to maintain the high efficiency of the original detector and reduce the complexity. Multiple sets of experiments on the NEU-DET dataset show that the improved YOLOv8n-CSG network model mAP@0.5 reaches 76.8%, compared with YOLOv8n, mAP@0.5 is improved by 6.9%, the accuracy is increased by 11.3%, the calculation cost is reduced by 37%, and the parameter quantity is reduced by 35.2%, showing a better detection ability for steel surface defects, and balancing the performance and complexity of the model.

Abstract:To address the limitations of traditional steel strip surface defect detection methods, such as insufficient feature extraction capability, restricted detection accuracy, and high computational resource consumption, this study proposes ESE-YOLO, a model based on YOLOv8, designed to effectively detect surface defects on steel strips. Firstly, to enhance the model’s ability to extract edge features, an EIEStem efficient front-end module is introduced. This module utilizes a SobelConv branch to extract edge information from images, combined with a pooling branch to capture essential spatial information, thereby improving the model’s perception of defect regions. Secondly, within the backbone network, shift-wise convolution is integrated with the C2f module to construct the C2f_SWC module. This integration expands the model’s field of view through shift operations, enhancing its ability to capture contextual information and further improving the accuracy of spatial feature extraction. Additionally, to optimize the structure of the feature pyramid network, the EMBSFPN module is employed. This module adaptively selects multi-scale convolutional kernels based on different feature layers, enabling progressive acquisition of multi-scale perceptual information. By weighted fusion of the importance of features across different scales, the detection accuracy is enhanced while significantly reducing the model’s parameter count and computational cost. Experimental results indicate that, compared to the original YOLOv8n, ESE-YOLO achieves a 4.1% improvement in mAP on the NEU-DET dataset, with a 26.8% reduction in parameters and a 64% decrease in floating-point operations. On the GC10-DET dataset, ESE-YOLO demonstrates a 9.9% improvement in mAP. Thus, ESE-YOLO significantly enhances detection accuracy while drastically reducing computational resource requirements, better meeting the deployment needs of resource-constrained devices in industrial scenarios.

Abstract:To improve the accuracy of deep learning in X-ray injection molding workpiece defect detection and realize higher precision nondestructive testing, an improved YOLOv8-seg internal defect segmentation model PIS-YOLO was proposed in this paper. Firstly, to reduce the number of parameters and improve the feature fusion capability, a multi-scale feature fusion and channel number reduction HG-Net module is designed in the backbone network to replace the traditional C2f module. The iRMB_EMA attentional fusion module is further introduced to enhance the deep transmission, and the feature fusion is completed by PAN-FPN with simplified redundant connections. Meanwhile, an additional output segmentation detection head is added to capture small defects, which improves the model’s accurate recognition of small target defects and defect edges. On the self-made data set of injection molding industrial parts, HG-Net module proposed in the backbone network section achieves a 22.03% reduction in computation under the same architecture compared with C2f module. On this basis, the overall precision of the model combined with the iRMB_EMA attention fusion module and additional output detection head is improved by 2.9% and 5.7%, respectively, compared with the benchmark model, and the model is lighter and less complex.

Abstract:Aiming at the problems of the current fan blade defect detection algorithm, such as insufficient detection accuracy, high incidence of false detection under complex background and inconvenient deployment of model, an improved YOLOv8n fan blade defect detection algorithm was proposed. Firstly, a new Extra-IB module and C2f-Extra-IB module are introduced to improve the key modules in MobilenetV2, which are used to reduce the number of model parameters to achieve lightweight and pass high-quality feature maps for subsequent feature fusion. Secondly, the AEMFP module is proposed to replace the SPPF module, which innovatively integrates the EMA attention mechanism and parallel substructure design to improve the multi-scale feature fusion and feature adaptive extraction capability of the algorithm. Finally, ELA attention mechanism is introduced into the neck network to reduce the influence of complex environment on the detection effect and improve the detection accuracy of small targets. Ablation experiments and comparison experiments were conducted using fan blade surface defect data set. The proposed algorithm mAP reached 81.7%, an increase of 5.1% compared with YOLOv8n. The number of model parameters and floating-point calculations were 2.09×106 and 5.4 GFLOPS, respectively, decreasing by 22.3% and 21.7%. The model size is reduced by 19.8% and the detection frame speed reaches 45.57 frames.It shows that the improvement measures proposed in this paper can not only improve the detection accuracy of the algorithm, but also achieve lightweight, which can meet the demand of using the detection equipment with limited computing resources such as UAV for efficient and accurate fan blade defect detection.

Abstract:Steel is the pillar industry of China, and its surface quality is the key to affecting the performance and price of steel. In order to solve the problems of poor accuracy, low efficiency and high model complexity in strip surface defect detection, a lightweight strip surface defect detection model (PGS-YOLO) was proposed and improved. Firstly, a more flexible PReLU activation function was introduced, and the slope of the negative region of the input data was adaptively adjusted through the learnable parameters, so as to improve the accuracy of the model to locate defects. Secondly, the Re-VGG is integrated into C3 to build a lightweight and efficient Re-C3 module to reduce the complexity of the model and improve the computational efficiency. Finally, the lightweight SCDown downsampling operation is adopted to reduce redundant calculations and improve the richness of feature fusion. Experimental results on the NEU-DET dataset show that the mAP of the model is increased by 6.7% to 79.9% compared with the benchmark model. The number of parameters and the amount of computation are reduced by 29.7% and 27.2%, respectively, and the FPS is increased by 2.7%, which better balances the relationship between detection accuracy, inference speed and lightweight. In addition, the model shows good generalization ability on both the WF10-DET dataset and the PCB_DATASET dataset, which meets the needs of actual engineering deployment and is expected to have important promotion and application value in engineering applications.

Abstract:With an aim to address the issues of low precision and susceptibility to background interference in steel surface defect detection, a YOLOv10n target detection algorithm based on fusion and shared parameters is proposed. Firstly, the backbone network incorporates the enhanced FasterNet lightweight network and the channel-first convolutional attention mechanism to enhance the capacity of the backbone network in representing multidimensional information. Secondly, the PCONV-C2F module is designed based on partial convolution (PConv) to tackle the problem of the disparity in the sensitivity field of the C2f module. Thirdly, wavelet pooling is utilized to address the problem of aliasing and background interference resulting from the up and down sampling mechanism in the original algorithm. Finally, a lightweight detection head is put forward to reduce the computational complexity of the model and enhance the accuracy of bounding box prediction by integrating shared parameters with dynamic distribution techniques. The mean average precision (mAP) mAP@0.5 of the improved algorithm on the NEU-DET dataset attains 86.3%, which is 8.1% higher than that of the original algorithm, and the precision reaches 86.8%, which is 18.7% higher than that of the original algorithm. The ablation and comparison experiments demonstrate that the improved algorithm exhibits excellent performance in the surface defect detection of steel and metal materials, which not only meets the requirement for efficient and accurate detection of steel surface defects in practical applications, but also significantly enhances the reliability and practicability of the detection.

Abstract:To address the challenge of small and highly variable defect sizes in photovoltaic panels, an improved YOLOv8n-based model named YOLO-RMFP is proposed. First, by integrating an efficient multi-scale attention mechanism with receptive field attention, a Receptive Field Mixed Attention mechanism is introduced. This mechanism enables the model to focus on features at multiple scales while addressing the parameter-sharing limitations of conventional multi-scale attention, thereby enhancing the detection accuracy for tiny defects in photovoltaic panels. Second, the Receptive Field Mixed Attention mechanism is integrated with the Spatial Pyramid Pooling module to enhance the model’s capability to capture multi-scale features and focus on complex regions. This integration improves the model’s ability to suppress noise in complex backgrounds, thereby further boosting the detection precision of small defects in photovoltaic panels. Then, feature maps of different resolutions from the YOLOv8n backbone are fused with an improved multi-scale feature fusion pyramid network. This enhances the interaction of feature information, enabling more comprehensive feature extraction and improving overall detection performance. Finally, based on the PIOU loss function, the model adjusts the weightings of defect samples according to their detection difficulty. This improves the localization accuracy and effectively mitigates the problem of sample imbalance in photovoltaic defect detection. Results from ablation and comparative experiments show that the YOLO-RMFP model improves detection accuracy, with mAP@0.5 and mAP@0.5:0.95 increasing by 3.1% and 6.5%, respectively. Precision and recall are also enhanced by 4.2% and 3.5%, respectively. These results demonstrate that the proposed model meets the performance requirements for photovoltaic panel defect detection.

Abstract:The prediction of satellite temperature telemetry data has important research and application value for monitoring satellite status and fault warning in ground operation and maintenance systems. However, traditional prediction methods have limitations such as low accuracy, insufficient robustness, and inability to provide probabilistic interval expressions. Therefore, this study proposes a multivariate temporal probabilistic prediction model TFM Diff for satellite temperature telemetry sequences. Firstly, a hybrid architecture based on gated recurrent units and discrete cosine transform was constructed to more accurately identify time-frequency domain dynamic patterns in telemetry data. Next, by integrating multi granularity features, complex modeling of short-term fluctuations and long-term trends in temperature telemetry data can be achieved, effectively analyzing the multi-scale characteristics of satellite temperature data. Finally, by combining the denoising diffusion model to comprehensively analyze the potential distribution patterns of the data, the probability interval expression of the prediction results can be achieved. Experimental verification based on four sets of real satellite temperature datasets shows that the continuous ranking probability score sum index for probabilistic prediction has improved the predictive performance of the proposed model by 6.26% to 27.77% compared to other mainstream methods, verifying its superior predictive performance, good applicability, and universality in space application scenarios.

Abstract:The accurate testing of electromagnetic parameters for foam absorptive materials presents a significant challenge. In this study, we developed an electromagnetic parameter (complex permittivity) characterization method based on the reflectivity calculation theory. By establishing a relationship between the electromagnetic parameters and reflectivity, we achieved precise characterization of the foam absorptive material’s electromagnetic properties. We propose a measurement system using the arcuate method, where the reflectivity for both vertical and horizontal polarizations is measured at different incident angles. The inversion problem is solved using the Newton-Raphson iteration algorithm. This approach only requires the amplitude of the reflectivity, thus eliminating the impact of phase errors. The proposed testing and characterization method only needs a single sample without the need for precise cutting, preserving the microstructure of the foam absorptive material and minimizing the influence of processing errors. To validate the proposed testing scheme and algorithm, experiments were conducted in the frequency range of 8 GHz to 12 GHz, and the results were compared with those obtained using the classic waveguide method (NRW). The experimental results demonstrate that the theoretical reflectivity values for expanded polypropylene (EPP) at different thicknesses and angles match well with the measured values. Except for resonance points, the deviation between the theoretical and measured values ranges from 0.087 dB to 0.806 dB, which meets the arcuate method testing requirement (±1 dB), thereby verifying the feasibility and effectiveness of the proposed electromagnetic parameter extraction algorithm.

Abstract:Instance segmentation is an important research direction in the field of computer vision, but the existence of the occlusion problem still prevents this task from being fully explored. To address the poor segmentation detection of occluded objects by current algorithms, which are prone to the problems of misdetection and omission, a novel duplex residual network （DRNet） is proposed, combining the EIoU occluded target segmentation model with the Mask R-CNN framework. First, DRNet is proposed to replace the original ResNet network, using fewer BN and ReLU layers to replace the traditional Conv-BN-ReLU structure, utilizing the conventional convolution and depth-separable convolution serial connection to enhance the image sensory field features, and mitigating the degradation problem of the network with the increase of the depth by the hopping connection. Second, the CEIoU NMS algorithm is used instead of the original NMS algorithm to effectively deal with the overlapping bounding box suppression problem through the clustering idea, and the introduction of the EIoU evaluation index increases the bounding box geometric information, which more accurately describes the degree of similarity between the bounding boxes, and reduces the network’s erroneous suppression of the bounding boxes of the occluded objects. Finally, the EIoU loss is used to replace the original Smooth L1 loss to accelerate the network convergence speed and improve the bounding box detection accuracy. In this paper, we first conduct pre-training on the public COCO 2017 dataset and experiments on different degrees of occlusion datasets, and the results show that compared with the original network, the proposed segmentation algorithm improves the Box AP and Mask AP by 1.7% and 1.3% on the COCO 2017 dataset, respectively; and both the bounding-box detection accuracy of the occluded object and the mask segmentation accuracy on the occlusion dataset are significantly improved on the occlusion dataset, confirming the effectiveness of the method for occluded object segmentation.

Abstract:To address the limitations of traditional vision-based methods in measuring the full dimensions of different workpieces, this paper proposes an online full-dimension inspection method for workpieces based on shape matching. The method inputs the target workpiece image into an improved Superpoint keypoint detection network to obtain all keypoints, which are then used to describe the workpiece contour. Then, the keypoint template and the keypoints of the target workpiece are fed into a point rendering layer. An enhanced Superglue feature matching algorithm with augmented keypoint location information is employed to achieve full matching, extracting keypoints that match the template points and measuring the distances between keypoints, thereby enabling full-dimension measurement of the workpiece. To validate the effectiveness of the proposed method, experiments were conducted, including gauge block size detection, calibration plate size detection, and electrochemical cell size detection. The experimental results indicate that for the size measurement experiment of a 25 mm Grade 0 gauge block (with an accuracy better than ±0.14 μm), the maximum deviation of the system’s ten repeated measurements was ±0.02 mm, and the standard deviation was 0.01 mm, demonstrating that the system has high repeatability accuracy. For the checkerboard calibration plate, the size measurement error does not exceed ±0.03 mm, verifying the feasibility of the proposed method. In the dimensional measurement experiment of primary batteries, the AAA battery size inspection had an error range of ±0.03 mm with an average processing time of 0.08 s, while the AA battery inspection showed an error of ±0.03 mm with an average time of 0.09 s. Both meet the enterprise’s production line requirements for online inspection, which demand ±0.05 mm accuracy and real-time detection within 0.1 s. Unlike traditional algorithms that require specific detection methods for different workpieces, the proposed approach exhibits strong adaptability to diverse dimensional detection requirements and is highly applicable for online full-size inspection of parts in industrial settings.

Abstract:The analysis of elements in drilling cuttings using laser-induced breakdown spectroscopy (LIBS) technology can not only provide information about underground formations, optimize the drilling process, but also enhance the safety and economy of drilling. Affected by the complex drilling environment, the LIBS of drilling cuttings generally exhibits a relatively severe baseline drift phenomenon, while the existing baseline correction methods are prone to issues such as baseline underestimation or overestimation. Therefore, an improved adaptive penalized least-squares baseline correction method is proposed. Based on the asymmetric penalized least-squares algorithm, the tanh function is introduced to automatically adjust the weight matrix according to the peak height of the spectral signal, and a smooth parameter automatic adjustment strategy is designed by utilizing the difference and standard deviation between the spectral data and the estimated baseline to balance the conflict between the smoothness and fidelity of the spectral data during baseline correction. Verification was conducted on both simulated spectra and the measured LIBS of drilling cuttings. The results indicate that the proposed method has lower root mean square error (RMSE) values on simulated spectra with different noise levels, and improves the quantitative analysis accuracy of elements on the measured LIBS of cuttings with the R2 values of 0.992 6, 0.993 0, 0.968 4, 0.969 1, and 0.977 4 for five elements, namely Si, Ca, Mg, Al, and Fe, respectively, all exceeding 0.96. It can effectively promote the element analysis of drilling cuttings in complex oil and gas environments using laser-induced breakdown spectroscopy technology.

Abstract:As the core component of automatic control system of hydraulic excavator, the reliability of hydraulic system pressure sensor directly affects the control performance of the whole excavator. To solve the key problem of loss of control system signal caused by pressure sensor failure under complex and severe working conditions, this study proposes a high-precision pressure data real-time prediction method based on depth learning. Firstly, based on the electro-hydraulic proportional control system of 37t hydraulic excavator, a test platform is established to collect the data of multi-source sensor under the actual excavation and loading operation condition; Secondly, the maximum information coefficient method is used to analyze the feature correlation, and the 125-dimensional original data is reduced to the 10-dimensional effective feature, and the high-quality data set is constructed by means of Kalman filtering and standardization; Then, the feature weight distribution module based on attention mechanism is designed, and the super-parameter configuration of long short term memory(LSTM) network is optimized by combining with sparrow search algorithm(SSA) to construct the SSA-LSTM-Attention fusion prediction model. Through the experimental verification of seven typical prediction models, such as convolutional neural network (CNN), gate recurrent unit (GRU) and LSTM, this method shows significant advantages in key pressure data prediction. The experimental results show that the mean absolute error and root mean square error of SSA-LSTM-Attention model are reduced by 54.45% and 54.56% respectively compared with the traditional LSTM model. The research proves that the proposed method can effectively solve the data compensation problem under sensor failure condition, and provide theoretical support for the fault tolerant design of intelligent control system of engineering machinery.

Abstract:In real surveillance scenarios, pedestrian re-identification tasks face numerous challenges, such as partial image occlusions (trees, people, cars, small objects, etc.) that lead to the loss of key information and a decline in recognition accuracy. To address issues like low recognition accuracy in occluded pedestrian re-identification tasks, methods that combine local and global features or use pose estimators are commonly employed. Although single-stream networks can achieve good recognition performance under partial occlusions, they fail to fully exploit the remaining critical feature information during processing. Therefore, we propose an occluded pedestrian re-identification method based on a multi-granularity dual-stream network. By designing a multi-granularity local feature extraction strategy, a dual-stream feature processing network, and a feature weight fusion module, the ability to extract key feature information is enhanced. This method employs a vision Transformer (ViT) to extract global features and divides them into multiple groups of local features. Subsequently, each group of local features is processed through a dual-stream feature processing network. The features obtained from the dual-stream network are then fused using a feature weight fusion mechanism, thereby more effectively mining key feature information. Experimental results on the Occluded-Duke, Market-1501, DukeMTMC-reID, and MSMT17 datasets demonstrate the effectiveness and validity of the proposed method, achieving mAP/Rank-1 indicators of 61.3%/68.3%, 89.0%/95.2%, 82.5%/91.1%, and 66.8%/84.5%, respectively.

Abstract:To address the technical challenge of measuring flame height in wildfire monitoring and risk early warning for transmission lines, this study proposes a method for measuring wildfire flame height by integrating the YOLOv9-SOEP algorithm with binocular stereo vision. Based on the YOLOv9 network architecture, the method introduces a small object enhancement pyramid (SOEP) module to construct an improved YOLOv9-SOEP target detection algorithm tailored for transmission line wildfire scenarios. To overcome the issue of weak texture features in flame images, a phase consistency method is adopted to achieve high-precision feature point extraction and matching in binocular wildfire images. Finally, a comprehensive flame height measurement model for transmission line wildfires is established through 3D coordinate transformation of feature points and pixel ratio calculation. Experimental results demonstrate that the improved YOLOv9-SOEP model achieves an average precision and recall of 85% and 89%, respectively, representing improvements of 4% and 19% over the original model, effectively addressing the issue of missed detection for small flame targets. The phase consistency-based stereo matching method effectively preserves the detailed features of flame targets in depth maps, achieving a matching accuracy of 92% while ensuring sufficient feature points. In simulated wildfire flame height measurement experiments, the measurement error was controlled within 6.48%. The proposed method provides a reliable solution for flame height measurement in transmission line wildfire monitoring and risk early warning.

Abstract:High-speed gas flow measurement, such as flare gas flow, is of great significance for energy saving and emission reduction, and for improving enterprise economic benefits. Ultrasonic flowmeter has a wide range, stable measurement, and is not affected by component changes. At present, ultrasonic flowmeter is the mainstream product of torch gas flow measurement. At present, this kind of products are basically foreign monopoly. The wave propagation time of domestic ultrasonic flowmeter fluctuates greatly in the measurement of high-speed gas, resulting in a large error of measurement results. In order to find a suitable algorithm for estimating the propagation time of sound waves, ultrasonic signals at gas flow rates ranging from 10 m/s to 70 m/s were collected from professional metering devices of instrument manufacturers. Four signal processing methods including envelope fitting method, autocorrelation method, cross-correlation method and wavelet time-frequency analysis method were studied. The experimental results show that wavelet time-frequency analysis has a good time of fly estimation effect. Standard deviation is smaller than 9 μs. But there are still fluctuations at high speeds. Averaging, least square method or secant slope method can eliminate data fluctuations and improve measurement accuracy and stability.