Abstract:This study proposes an infrared defect detection and evaluation framework for hypervelocity impacts based on virtual objective vectors. The framework integrates multi-objective optimization with infrared feature extraction, employing a clustering multi-objective optimization approach to extract infrared features for different defect categories. Dynamic time warping is used to extract typical transient thermal responses, while virtual objective vectors are applied to extend the weight vector set and enhance the estimation accuracy of the Pareto front. Experimental results demonstrate that this method effectively improves infrared defect detection accuracy under hypervelocity impacts, providing a reliable basis for spacecraft damage assessment.

Abstract:Alzheimer′s disease (AD) is a neurological disorder that primarily affects a person′s brain cells and is the main form of dementia; due to its irreversible nature, early diagnosis is critical to slowing the progression of the disease. Structural magnetic resonance imaging (sMRI) and fluorodeoxyglucose positron emission tomography (FDG-PET) are two imaging techniques that are widely used in neurodegenerative disease research, and combining these two images to assess the brain state can improve the accuracy of the results. In this paper, we propose a multimodal fusion framework based on Vision Transformer, which extracts features from unimodal images through a self-attentive vision transformer, and at the same time focuses on the similarity of the features of the two images by using an interactive attentional fusion network, which strengthens the independent characterization ability of each modality, and also improves the interactivity of the two modalities. At the same time, a deep confidence network is used to reduce the redundancy of the extracted features and improve the complementary information of different modalities, and finally an integrated classifier is used to make AD classification results. The ADNI dataset is selected and the classification performance of the proposed network is evaluated, and the accuracy, sensitivity and specificity reach 94.65%, 93.24% and 95.62%, respectively, and the proposed method achieves superior results in the AD classification task compared to current fusion methods.

Abstract:In the broadband signal sampling system, the sampler is an important component. It adopts a symmetrical structure internally to reduce common-mode interference during sampling. It achieves transient sampling and holding of broadband signals by triggering the internal fast recovery diode to conduct rapidly with a narrow pulse signal. However, inconsistent characteristics of the internal diodes and the width of the external narrowband trigger signal being too wide can both lead to poor symmetry of the two output signals and inaccurate sampling. To address these issues, an offset adjustment circuit based on FPGA was designed to compensate for the offset of the sampler′s output signals. After offset compensation, the amplitude error of the two output signals is approximately 3 mV, meeting the symmetry requirements. A narrow pulse trigger circuit was designed using step fast recovery diodes. The trigger signal generated by the trigger circuit has a falling edge time of about 50 ps and an amplitude of 17 V. The measured results show that this signal can trigger the sampler to achieve sampling of signals up to 30 GHz.

Abstract:To address the issues of false detections and missed detections in underwater fish recognition under complex environmental conditions, this paper proposes FD-YOLO, a recognition method based on YOLOv8 incorporating dynamic adaptive feature learning. First, in the backbone, we design a MRFA module that combines parallel multi-scale convolution with RFA to enhance the network′s ability to capture fine-grained differences in fish features. Second, in the neck, we introduce the ECARU module, which integrates a dual-channel fusion structure with the CARAFE mechanism. This upsampling module adaptively reweights features to improve the reconstruction of local fish details. Finally, to mitigate recognition bias caused by class imbalance, we adopt Varifocal Loss, which includes a dynamic adjustment factor, to improve the accuracy of underwater fish localization. Experimental results demonstrate that, compared with YOLOv8n, the proposed FD-YOLO achieves improvements of 3.6%, 4.9% and 3.7% in Precision, Recall and mAP50.95, respectively. Moreover, the parameter size and computational cost are reduced to 2.5 MB and 6.8 GFLOPs. These findings demonstrate that the proposed method can provide technical support and reference for automated detection and monitoring of underwater targets.

Abstract:This paper addresses the challenges of segmenting rural roads in remote sensing images, including small pixel proportion, irregular shapes, shadow occlusions, and blurred edges. To improve the segmentation accuracy of small and single-object rural roads, we propose an improved DeepLabV3+ semantic segmentation model. We employ MobileNetV3 as the backbone for parameter reduction and enhanced accuracy. A global attention mechanism is incorporated to improve global information extraction and generalization. Depthwise separable convolutions replace standard convolutions in the ASPP module to minimize information loss and computational cost. Experiments on a self-built satellite remote sensing road image dataset demonstrate significant improvements, achieving an MIoU of 84.45% and MPA of 92.32%, outperforming the original DeepLabV3+ by 4.63% and 6.48%, respectively, with a parameter size of only 6.30×106. Validation on the public CHN6-CUG dataset confirms the model′s effectiveness, showing MIoU and MPA improvements of 3.05% and 5.54% to reach 79.64% and 88.13%, respectively. These results indicate that our lightweight, improved model effectively enhances rural road segmentation accuracy and efficiency.

Abstract:In response to the demand for the intelligence of ammunition systems and the integration of onboard computers, a design scheme for an integrated onboard computer architecture based on intelligent processors has been proposed. The integrated onboard computer can utilize intelligent modules to perform smart recognition of image data transmitted from the guidance head and can control various systems on the projectile through an information processing module. This design makes full use of various processor resources, creating a robust hardware architecture support platform for high-speed real-time large-capacity data processing, complex algorithms, and data transmission on the projectile. By employing a parallel software design strategy within a layered architecture, system computation tasks are evenly mapped to different processors, thereby achieving the goal of multi-task real-time parallel processing. The final prototype test results indicate that the integrated architecture design is reasonable, carrying significant military implications and engineering application value.

Abstract:For surface defect detection of carbon fiber reinforced polymer (CFRP) tow placement, where line-structured light images suffer from high-intensity noise regions and difficulties in extracting light stripe centers at fine gaps, this paper proposes an adaptive fast line-structured light center extraction algorithm. The method first reduces image complexity through preprocessing and light stripe ROI extraction. Subsequently, a contour tracking algorithm eliminates high-intensity noise and determines light stripe boundaries. Finally, an 8th-order filter convolution operator extracts skeleton points, combined with a normal centroid method to calculate sub-pixel center coordinates. When processing multiple defective tow placement images, our algorithm achieves optimal point cloud reconstruction accuracy. It requires only 5.36 s to extract center lines from 50 images, with a standard deviation of merely 0.38 pixel under extreme lighting conditions. Experimental results demonstrate that the proposed algorithm maintains accurate light stripe center extraction despite high-intensity noise and complex defect interference, exhibiting high precision, efficiency, and robustness. This provides reliable technical support for automated defect detection in composite material layup processes.

Abstract:Synthetic Aperture Radar and optical images capture surface features through distinct dimensions, providing highly complementary information for land classification research with significant application value. However, existing MCANet-CM algorithms struggle to effectively capture target contours in multimodal data during cross-modal feature interaction, resulting in insufficient spatial detail representation of fused features for object boundaries in complex scenarios. This makes the effective integration of dual-modal data for precise pixel-level classification remain a critical challenge. To address this issue, this paper proposes an enhanced multimodal remote sensing image semantic segmentation algorithm based on improved MCANet-CM. The algorithm introduces the DyCPCA attention mechanism, which dynamically calibrates inter-channel dependencies to adaptively enhance feature responses related to target contours, thereby significantly improving the model′s capability to capture fine-grained information from multimodal data. Simultaneously, a Rectangular Self-Calibration Module is incorporated, which constructs an asymmetric receptive field structure to strengthen the model′s perception of edge information across different orientations, markedly enhancing localization accuracy for foreground objects. Through the synergistic operation of these two modules, effective fusion of optical and SAR data is achieved. Experiments on the WHU-OPT-SAR dataset demonstrate that compared with the baseline MCANet-CM model, the improved model achieves 2.85% and 2.81% enhancements in mean Intersection over Union and mean F1-score, respectively. When compared with state-of-the-art algorithms like FTransUNet, the proposed model also exhibits superior segmentation performance.

Abstract:Aiming at the problems of industrial robots in torsional vibration optimization, such as the tendency to fall into local optimum, slow optimization speed, and poor optimization effect, this paper proposes an improved method based on the non-dominated Sorting Whale Optimization Algorithm (NSWOA). Firstly, by introducing the non-dominated sorting algorithm to perform Pareto optimization on the three objectives, the exploration ability of the solution space and the distribution performance in multi-objective optimization are significantly enhanced. Secondly, the NSWOA is combined with input shaper technology. Through transfer function transformation, online signal acquisition and offline optimization processing are realized, which avoids the problem that online optimization is prone to exciting system vibration, while offline modeling has low accuracy. The algorithm is verified on the B&R test platform. The results show that compared with PSO, DBO and ACO, the non-dominated sorting whale optimization algorithm based on the input shaper shows significant advantages. The overshoot is reduced by 80.6%, 92.1% and 92.8%, respectively. The system adjustment time is 10.9%, 7.2% and 6.7% of the other three methods, respectively. While significantly suppressing the system torsional vibration, the dynamic performance of the system is only slightly sacrificed, achieving a fast and vibration-free system response. This verifies the rationality and superiority of the algorithm.

Abstract:A proactive active and reactive power collaborative optimization strategy considering low-carbon demand is proposed to address issues such as voltage exceeding limits caused by the high proportion of new energy connected to the distribution network. Firstly, in order to fully tap into the potential of carbon reduction in the system, a tiered carbon trading model is established to stimulate load side adjustment of electricity consumption behavior and achieve low-carbon response. Then, considering the operational requirements of the distribution network, a collaborative optimization model is constructed with the goal of minimizing network loss, voltage deviation, and comprehensive operating costs, and compensating equipment and flexible loads as decision variables. To overcome the drawbacks of slow convergence speed and susceptibility to local optima in the Pelican algorithm, an improved Pelican algorithm is proposed. In the early stage of the algorithm, Bernoulli chaotic mapping is used to initialize the population and sparrow vigilance mechanism and nonlinear inertia weights are introduced to balance and enhance the exploration and development capabilities of the algorithm. In the later stage of iteration, the Cauchy perturbation is used to enhance the algorithm′s ability to escape from local optima. Finally, the effectiveness of the proposed strategy and algorithm was verified through simulation of an improved IEEE33 node system.

Abstract:The accuracy of wind power prediction is crucial for ensuring the sustainable and stable operation of power grids. To address the issue of inadequate prediction accuracy caused by the volatility and stochasticity of wind power data, this study proposes a decomposition-prediction model based on the Successive Variational Mode Decomposition (SVMD) algorithm for data decomposition, combined with a Bidirectional Temporal Convolutional Network (BiTCN) and Bidirectional Long Short-Term Memory Network (BiLSTM) for prediction. The Splendid Fairy-wren Optimization Algorithm enhanced with Newtonian method (SFOA-N) is employed to optimize SVMD′s penalty factor and the hyperparameters of the prediction model, thereby improving local search capability. To resolve the technical challenge that the exponentially growing dilation rate in BiTCN struggles to adapt to complex patterns across different time series, an innovative dynamic dilation rate prediction module is proposed. This module automatically adjusts dilation rates according to varying segments of input data, significantly enhancing prediction performance. Experimental results demonstrate that compared with standalone BiTCN models, the optimized SVMD-IBiTCN-BiLSTM model achieves a coefficient of determination of 0.998 2, with mean absolute percentage error, root mean square error, and mean absolute Error reduced by 3.57, 9.94, and 7.21 respectively, demonstrating superior forecasting accuracy.

Abstract:A multi-strategy based golden jackal optimization algorithm is proposed to address the problems of poor population quality, slow convergence speed and easy to fall into local extremes faced by the golden jackal optimization algorithm in solving constrained optimization problems. First, in order to increase the diversity of the population and improve the quality of the initial solution, a chaotic elite collaborative initialization strategy is used to generate an elite population; then， an energy regulation mechanism is introduced to coordinate the global search and local optimization; finally, a fusion mutation method is designed for the individual differences in the population in order to prevent the problem of local extremes. The improved algorithm is proved to have better convergence performance and faster convergence speed through the comparison test of standard test functions. In addition, experiments on the CEC2021 test function and the pressure vessel design optimization problem further demonstrate the effectiveness and practicality of the improved golden jackal optimization algorithm in single-objective constraints and multi-objective constraints problems through convergence analysis, robustness test, and validation of Wilcoxon′s rank sum statistics.

Abstract:Individualized intravenous drug therapy imposes higher demands on the rapid and precise detection of drug components, especially as the identification of trace substances in complex matrices remains challenging. This study proposes a frequency-encoded, multi-wavelength detection system based on near-infrared spectroscopy, which mitigates water-induced spectral overlap and overcomes the limitations of single-wavelength lasers in solute identification and quantification. The system independently encodes eight groups of LD lasers within the range of 850 nm to 1 550 nm, and combines phase-locked amplifier algorithms for signal demodulation, effectively suppressing interference between wavelengths and improving signal accuracy and response speed. It integrates a high-sensitivity InGaAs array detector, capable of capturing the absorption features of C—N (1 380~1 430 nm) and N—H (1 500~2 100 nm). By establishing a spectral-to-concentration mapping model using neural networks, the detection time is reduced to less than 2 seconds, significantly improving efficiency. Experimental results show that in hydrochloride ambroxol (0.1~2 mg/mL) and biapenem (1~5 mg/mL) solutions, the concentration detection error is controlled at ≤5% (n=20), outperforming traditional HPLC methods (5%~8%). These results demonstrate the system′s potential in rapid drug component detection and individualized drug monitoring, with broad application prospects.

Abstract:In order to research the time-frequency characteristics and energy distribution of explosion vibration signals, a HHT based feature extraction method of explosion vibration signal was constructed in this paper. And we used the method to analyze the vibration velocity signals of a certain explosion experiment. The results showed that denoising the high-frequency IMF components after EMD decomposition using wavelet threshold method could effectively remove high-frequency noise from the signal and improve the accuracy of Hilbert spectrum analysis. Through comparative analysis of the Hilbert time-frequency spectrum, marginal spectrum and instantaneous energy spectrum of the signals at each measuring point, the instantaneous relationships among time-frequency-energy and variation pattern with explosion center distance were obtained. Within a range of 80 meters from the explosion center, the vibration signal energy is mainly distributed in the 0~250 ms time range and 0~150 Hz frequency range. With the increase of distance from explosion center, the starting time, peak speed time and energy concentration period of the signals at each measuring point was gradually delayed. In addition, the amplitude of vibration, instantaneous energy and high-frequency components in the signals were gradually decayed.

Abstract:To address the issues of low power capture efficiency, slow dynamic response, and weak interference resistance faced by direct-drive wave energy conversion devices in complex sea conditions, this paper proposes a control algorithm that combines neural networks with model prediction. By enhancing system robustness through a high-precision wave excitation force prediction model and combining it with a rolling optimization algorithm under multi-objective constraints, the device achieves maximum power generation under irregular wave conditions. First, a three-stage fusion prediction model with spatio-temporal feature decoupling capability is constructed. Compared to traditional models, this model reduces the mean squared error and mean absolute error of irregular wave excitation force prediction by 39.96% and 63.39%, respectively, with a temporal fitting accuracy of 98.9%. The prediction model is then embedded into a rolling optimization framework, where high-precision irregular wave excitation force predictions provide feedforward disturbance compensation for control, aligning motor current with the current at maximum power, thereby achieving the goal of maximizing power generation. Experiments demonstrate that the improved model predictive control achieves significant breakthroughs compared to the traditional autoregressive integral moving average model method under two irregular wave conditions (JS and PM) with wave heights of 0.3~0.6 m and periods of 3~6 s: average power increases by 50%~141%, and cumulative energy increases by 38%~189%, validating the significant advantages of the proposed method in enhancing the comprehensive performance and dynamic robustness of direct-drive wave energy conversion systems.

Abstract:Aimed at the long-term stability and reliability problem for ultrasonic water meters, this paper first analyzes the possible reasons caused the problem from measurement principle, and proposes new technology programs from measurement principle, measurability design, and measurement equipment. Firstly, it proposes a measurement principle of ultrasonic flight time based on echo attenuation detection. It avoids the first echo detection error caused by hardware attenuation or noise, which exists in the measurement principle of ultrasonic flight time based on first echo detection. It also proves the feasibility of the echo attenuation detection-based measurement principle. Secondly, it proposes the design method and equipment of ultrasonic liquid flow instrument with improved measurability. Thirdly, it proposes a test system suitable for ultrasonic liquid flow instrument, to resolve the problems in current ultrasonic water meter measurement. By providing systematic technology programs for the long-term stability and reliability problem, this paper lays a good foundation for the commercial application of ultrasonic water meters. The present analysis, principle and method can also be applied to other ultrasonic liquid flow meters such as ultrasonic gas meters and ultrasonic oil meters.

Abstract:Aiming at the problems of low detection accuracy and large model size existing in the current foreign object detection algorithms for railway catenaries, this study proposes a foreign object detection algorithm for catenaries (FRDW-YOLOv8) based on the improved YOLOv8. Firstly, we propose the integration of a FasterNet module into the backbone network to construct the C2f-Faster module, which effectively reduces model complexity and enhances computational efficiency. Secondly, the Receptive-Field Coordinate Attention mechanism (RFCA) is introduced in the feature extraction stage to increase the model′s attention to the foreign object areas of the catenary and allocate more attention to them. Then, a dynamic upsampler (Dysample) is adopted in the neck network, which can retain more detailed information of the foreign objects on the catenary. Finally, the WIoU v3 loss function is used to improve the overall performance of the detection model by dynamically adjusting the weight factors. The experimental results show that the mAP value of the improved algorithm reaches 95.1%, which is 2.8% higher than that of the YOLOv8 model, and the floating-point operations and the number of parameters of the model are only 7.3 G and 2.7 M respectively. The improved algorithm further improves the detection accuracy of the model and makes the model lightweight. It fully demonstrates that the detection performance of the improved algorithm is superior to the current mainstream algorithms and can better complete the task of detecting foreign objects on railway catenaries.

Abstract:As the core link of intelligent manufacturing quality control, the detection accuracy and real-time performance of surface defects in industrial products are crucial for industrial production. Aiming at the key problems of insufficient local feature sensitivity and high computational redundancy faced by existing unsupervised anomaly detection methods in complex industrial scenarios, an improved multi-scale feature fusion detection algorithm based on PatchCore is proposed. Firstly, by introducing a multi-scale feature fusion processing method with self attention mechanism, the layer 3 feature map is fused with self attention mechanism and average pooling to enhance the algorithm′s ability to capture local and global abnormal features; propose a channel aggregation dimensionality reduction method, which randomly divides the original features into several continuous subgroups and aggregates each group of features to generate low dimensional features, achieving the goal of reducing computational redundancy while preserving some local information of the original features; build transfer learning models to enhance the algorithm′s generalization ability in anomaly detection tasks and improve the detection accuracy of actual industrial projects. Through defect detection experiments on memory heat sink images, the results show that the improved algorithm improves AUROC by 2.28% and F1Score by 4.89% compared to the original algorithm, which can meet the requirements of high efficiency and high precision in industrial scenarios.

Abstract:Road damage increases the likelihood of traffic accidents, posing a serious threat to traffic safety. Therefore, real-time monitoring of road conditions is crucial for ensuring road safety and effectively managing infrastructure. To address the issues of insufficient detection accuracy and small target detection challenges in existing road defect detection methods, this paper proposes an improved RT-DETR-based road defect detection algorithm. First, partial convolution (PConv) is introduced to reconstruct the RT-DETR backbone network, effectively reducing computational overhead. Second, a triplet attention mechanism is integrated into the backbone network to enhance the model′s sensitivity to multi-dimensional features, enabling more precise capture of image details. Next, a BiFPN-based feature pyramid network is employed to optimize the CCFM feature fusion module, and S2 features are introduced to improve the detection performance of small targets. Finally, the DySample upsampling operator is utilized to capture more local details and semantic information, further enhancing the model′s ability to detect small targets. Experimental results show that the improved algorithm achieves a 3.6% increase in mAP@50 on the RDD2022 dataset compared to the original RT-DETR model, with a 12.5% reduction in the number of parameters and a detection speed of 66 fps. Compared with other object detection algorithms, the improved algorithm demonstrates significant advantages in both detection accuracy and speed, making it more suitable for practical applications in road defect detection.

Abstract:Road disease detection is crucial for traffic safety and road maintenance, but existing algorithms generally suffer from low detection accuracy, high computational costs, and difficulty in deploying on mobile devices. To address these issues, we propose a lightweight multi-scale road disease detection algorithm LMR-YOLO-P based on YOLOv8n. By designing a multi-scale group conv module to adapt to the variable sizes of road diseases, and constructing a light weight shared detection head to reduce computational costs while preserving fine details, introducing receptive field attention convolution RFAConv to enhance global information capture capability, combining the DFP module and efficient local attention mechanism to build a SAC module for enhanced multi-scale feature fusion, and finally utilizing the layer-adaptive sparsity for the magnitude-based pruning method to further compress the model. Experimental results show that on the RDD2022 dataset, the algorithm improved mAP50 by 1.8% compared to the YOLOv8n network, while reducing parameter count and computational cost by 46% and 40% respectively, successfully achieving lightweight and real-time high-precision detection of road diseases, providing an effective tool for intelligent road maintenance.

Abstract:Aiming at the problems of limited adaptability, loss of details and unclear features faced by target detection in low-light environment, the edge-driven detection method ED_YOLO is proposed. Firstly, the HESM module is proposed to extract edge information through the Sobel operator, guide the interaction of multiple features, and improve the sensitivity of effective information. Secondly, the C2f_DRM module is designed to efficiently integrate local details and global context information. Then, the LFAM module is constructed. Based on shared convolution, the adaptive control method of features of different scales is optimized to effectively reduce the loss of detail information. Finally, the RepGFPN module is introduced to improve the multiscale feature extraction capability of the model by using reparameterization technology. Experimental results on the ExDark dataset show that the mAP50 of the proposed method reaches 72.17%, which is 2.87% higher than the original YOLOv8n, achieving better detection effect.

Abstract:Aiming at the problems of low detection accuracy, large model parameters, and poor real-time performance of existing algorithms in surface defect detection of rolling bearings, an improved YOLOv10n rolling bearing surface defect detection algorithm is proposed. On the backbone network, redesign C2f using GhostConv, MSMHSA module, and CGLU module, construct CGMC2f module to enhance the model′s feature extraction capability and reduce the model′s parameter count; in SPPF, the SPPF-LSKA module is designed by combining GroupConv, Residual-Conv, and Fusion-Conv modules to construct a new GRFSPPF-LSKA module, effectively solving the problem of information loss and improving the model′s multi-scale feature extraction and fusion capabilities; on the Neck network, combining the multi-scale feature weighted fusion of BIFPN, MAF-YOLO network, and EMCAD module, an EMBS-FPN network is constructed to improve the detection accuracy of the model, reduce the number of model parameters, and make the model lightweight; drawing on the Focal-loss approach, optimize the CIoU loss function to Focaler-CIoU to accelerate the convergence speed of the model. The experimental results showed that the improved YOLOv10n achieved a mAP of 92.6%, an increase of 2.7% compared to the original model, a reduction of 0.45 M in parameter count, and a decrease of 0.6 GFLOPs in computational complexity, better meeting the real-time detection requirements of rolling bearing surface defects.

Abstract:Pavement disease detection is particularly important in road maintenance, and the YOLOX-GED algorithm is proposed for the problems of complex background of pavement image and large difference of disease scale in pavement disease detection. On the basis of YOLOX-s algorithm, the algorithm firstly designs CSP_Ghost module to replace CSPLayer module, which reduces the number of network parameters and at the same time strengthens the feature extraction ability of the network; secondly, introduces the ECA attention mechanism, which strengthens the feature fusion effect of the network, and improves the recognition accuracy of the network on the pavement lesions; lastly, designs the pyramid structure of DSPPF space, which increases the diversity of features and strengthens the recognition accuracy of the network on the pavement lesions. that increases the diversity of features and strengthens the extraction and fusion of multi-scale contextual information. Experiments on the RDD2020 dataset show that the mAP of the YOLOX-GED algorithm is 5.32% higher than that of the YOLOX-s algorithm, and at the same time, the amount of model parameters is reduced by 7.9%, which makes it easier to deploy to mobile devices.

Abstract:Anomaly detection aims to identify abnormal patterns in data, and it has significant application value in various fields such as industrial inspection. The current mainstream anomaly detection methods employ unsupervised models such as auto-encoder. These models use fully connected layers or convolutional layers for the data processing during encoding and decoding, which can lead to problems such as lack of interpretability and semantic errors. This paper proposes a combination method of the semi-nonnegative matrix factorization model and network training to design a semi-nonnegative matrix factorization neural network for anomaly detection. Due to the characteristic of the semi-nonnegative matrix factorization model that “local superposition constitutes the whole”, this network can better preserve semantic information and is also interpretable. Additionally, the feature matrix of this network is updated as weights during the training of the network, which effectively solves the problem of local optimal solutions existed in the traditional semi-nonnegative matrix factorization model. The anomaly detection performance of this network was tested on three datasets. The experiment results show that it outperforms mainstream auto-encoder and variational auto-encoder methods by more than 3 percentage points in continuous data, and achieves comparable results in discrete data. Compared with the detection method based on the traditional semi-nonnegative matrix factorization model, this network has significantly improved in all detection metrics, with the highest improvement reaching 12%. This network is a beneficial exploration that utilizes traditional matrix factorization model to construct neural network, and it can effectively solve the anomaly detection problem.