GIS internal foreign matter defect monitoring method based on improved YOLOv3

Abstract

The invention discloses a GIS internal foreign matter defect monitoring method based on improved YOLOv3. The GIS internal foreign matter defect monitoring method comprises the steps that firstly, signals are collected, wherein the signals in high-voltage equipment are collected through a sensor; then surveying and mapping a PRPD graph: generating the PRPD graph according to the collected signals; gray scale identification: reading the content in the PRPD image by using a gray scale identification algorithm; and finally, classifying by a CM-YOLOv3 algorithm: identifying the type of the foreign matter in the GIS cavity according to the read content. According to the invention, real-time monitoring and positioning of foreign matters in the GIS equipment can be realized, and personal errors and potential safety hazards caused by traditional manual inspection are effectively avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-voltage equipment monitoring, and in particular relates to a GIS internal foreign body defect monitoring method based on improved YOLOv3. Background Art

[0002] GIS internal foreign body defects refer to the presence of foreign matter inside the GIS equipment that should not exist and affects the performance and function of the equipment. These foreign matter can be any material that is not a normal part of the equipment, such as impurities, debris, metal chips, etc.

[0003] Foreign matter defects inside GIS may cause various adverse effects on the operation of the equipment, such as causing arc, short circuit, open circuit and other faults, reducing the insulation performance of the equipment, causing unstable electrical performance, etc. These problems may affect the reliability, safety and stability of the equipment.

[0004] The causes of foreign matter defects inside GIS can be various factors during manufacturing, transportation, installation, use and maintenance. For example, there may be impurities or substandard materials in the materials used in the manufacturing process; during transportation and installation, necessary protective measures are not taken, resulting in foreign matter entering the equipment; during use and maintenance, improper operation, reduced equipment sealing performance, incomplete maintenance and cleaning, etc. may all lead to foreign matter entering the equipment.

[0005] As a common form of failure in high-voltage equipment, foreign body defects in GIS may cause equipment damage or accidents if not discovered and resolved in time. Effective monitoring and identification of foreign body defects in GIS are the key to ensuring the safe operation of equipment. At present, some monitoring methods mainly use the collected signals for analysis, such as PRPD diagrams, but how to accurately identify the type of foreign body defects in GIS is still challenging. Summary of the invention

[0006] The purpose of the present invention is to provide a GIS internal foreign body defect monitoring method based on improved YOLOv3, which can realize real-time monitoring and positioning of foreign bodies inside GIS equipment, and effectively avoid human errors and safety hazards caused by traditional manual inspection.

[0007] The technical solution adopted by the present invention is a GIS internal foreign body defect monitoring method based on improved YOLOv3, which is characterized by being implemented specifically according to the following steps:

[0008] Step 1: Collect signals: collect signals from high-voltage equipment through sensors;

[0009] Step 2, mapping PRPD map: generating PRPD map based on the collected signals;

[0010] Step 3, grayscale recognition: use grayscale recognition algorithm to read the content in the PRPD image;

[0011] Step 4: CM-YOLOv3 algorithm classification: Identify the type of foreign matter inside the GIS cavity based on the read content.

[0012] The present invention is also characterized in that:

[0013] Step 2 is implemented as follows:

[0014] Step 2.1, signal preprocessing: Before generating the PRPD graph, the collected signal data is preprocessed, including signal amplification, signal filtering, and signal denoising;

[0015] Step 2.2, PRPD diagram generation: After the signal preprocessing is completed, the PRPD diagram is generated using MATLAB software. The PRPD diagram uses the peak amplitude and phase of the discharge pulse as coordinate axes, and each partial discharge pulse corresponds to a point in the diagram;

[0016] Step 2.3, data analysis: Analyze the generated PRPD map, and identify the partial discharge signals of different types of foreign objects inside the GIS by observing the aggregation and distribution patterns in different areas;

[0017] Step 2.4, identification and classification: The generated PRPD diagram intuitively displays the characteristics of partial discharge signals in GIS equipment, providing an important reference for fault diagnosis and maintenance work; in the PRPD diagram, the partial discharge signals generated by different types of foreign matter will show different characteristics. By observing these characteristics, different types of partial discharges can be identified and classified, as follows:

[0018] In the PRPD image, the characteristic points are concentrated in the narrow phase range (0°-30°) and (180°-210°). After identifying these points, it is confirmed to be bubble discharge;

[0019] In the PRPD diagram, the characteristic points are distributed in a wide phase range (30°-90°) and (210°-270°), and after identifying these points, it is confirmed to be a contaminated discharge;

[0020] In the PRPD graph, the characteristic points are distributed in the entire phase range, i.e., the narrow phase range (0°-360°), and after identifying these points, it is confirmed to be a tip discharge;

[0021] In the PRPD diagram, the characteristic points are distributed in the entire phase range, i.e., the wide phase range (0°-360°). After identifying these points, it is confirmed that it is a suspended metal discharge;

[0022] By performing trend analysis on the PRPD graph, changes in signal strength and frequency of occurrence can be discovered, and the working status of the equipment and the development trend of potential faults can be determined.

[0023] Step 3 is implemented as follows:

[0024] Step 3.1, read PRPD image: load PRPD image into MATLAB, and use MATLAB's image processing to perform operations: including image enhancement, feature extraction, and target detection. By processing the PRPD image, useful information can be extracted to understand the characteristics of the partial discharge signal inside the equipment. In addition, trend analysis using MATLAB can intuitively display the data changes in the PRPD image, help to promptly discover potential problems and take corresponding preventive measures.

[0025] Step 3.2, data preprocessing: Before grayscale recognition, the image needs to be preprocessed;

[0026] Step 3.3, feature extraction: determine the features you want to extract from the PRPD image. For partial discharge identification, the features include the number of groups, amplitude distribution, and phase distribution.

[0027] Step 3.4, grayscale recognition: Use grayscale recognition algorithm to classify and identify features;

[0028] Step 3.5: Result analysis: Analyze the classification results and evaluate the performance of the algorithm.

[0029] Step 4.1: Foreign body size clustering is as follows:

[0030] The new distance formula (1) is used to replace the Euclidean clustering. Box represents the clustering result, and centroid represents the cluster center:

[0031] d(box,centroid)=1-10U(bpx,centroid) (1)

[0032] By clustering analysis of foreign body samples, it is found that the average intersection-union ratio gradually increases with the increase of the number of anchor frames. When the number of anchor frames is greater than 9, the average intersection-union ratio increases slowly. Therefore, the number of anchor frames is selected as 9, and K-means clustering is used to obtain the best nine groups of anchor frame values, which are: (10.32), (13.14), (18.22), (19.41), (22, 35), (35, 59), (56, 74), (65, 98), (82, 54);

[0033] Step 4.2, Improved backbone network CSP-Darknet36:

[0034] The CSPNet structure is applied to the original backbone network Darknet-53, and the network model parameters are reduced by reducing the number of convolutional layer channels; the gradient propagation path is increased and the gradient splitting and merging strategy is used to reduce the reuse of gradient information and increase the learning ability of the convolutional neural network;

[0035] The Darknet53 network has a total of 5 residual blocks, each of which has 1, 2, 88, and 4 residual units respectively. The residual units of the relatively redundant 3rd, 4th, and 5th residual blocks are deleted to 4, 4, and 2 to form an improved feature extraction network called CSP-Darknet36;

[0036] Step 4.3, improved feature pyramid MFPN:

[0037] Improvements are made on the original feature pyramid by fusing three-scale feature information. The feature maps of 104*104, 52*52, and 26*26 scales are first downsampled, and then fused with the 52*52, 26*26, and 13*13 feature maps before the output layer. Finally, they are upsampled with the 26*26 and 13*13 feature maps and combined to output feature information of three scales.

[0038] The beneficial effect of the present invention is that the GIS internal foreign body defect monitoring method based on the improved YOLOv3 can generate a PRPD map by subsequent processing of the collected signal and use the grayscale recognition algorithm to perform feature recognition to accurately identify different types of partial discharges generated by different foreign bodies. In this way, the monitoring effect of foreign bodies inside the GIS cavity can be effectively improved, and potential faults can be discovered in time. The reliability and accuracy of this method make it widely used in the industrial field and play an important role in equipment management and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a PRPD map generated in the GIS internal foreign body defect monitoring method based on the improved YOLOv3 of the present invention;

[0040] Figure 2 It is a structural diagram of CM-YOLOv3 in the GIS internal foreign body defect monitoring method based on improved YOLOv3 of the present invention;

[0041] Figure 3 It is a K-means clustering result diagram in the GIS internal foreign body defect monitoring method based on improved YOLOv3 of the present invention;

[0042] Figure 4 It is a CSPNet structure diagram in the GIS internal foreign body defect monitoring method based on improved YOLOv3 of the present invention;

[0043] Figure 5 It is a CSP-Darknet36 network structure diagram in the GIS internal foreign body defect monitoring method based on improved YOLOv3 of the present invention;

[0044] Figure 6 This is a comparison diagram of the FPN and MFPN structures in the GIS internal foreign body defect monitoring method based on the improved YOLOv3 of the present invention. DETAILED DESCRIPTION

[0045] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] The present invention is based on the improved YOLOv3 GIS internal foreign body defect monitoring method, which is specifically implemented according to the following steps:

[0047] Combination Figure 1 ,Step 1, collecting signals: collecting signals from high-voltage equipment through sensors;

[0048] Step 1 is as follows:

[0049] In order to improve the performance of the sensor and the accuracy of signal acquisition, the collected signals can also be pre-processed by combining technologies such as signal amplification, filtering and denoising. By comprehensively utilizing high-quality sensors and optimized signal processing methods, the accuracy and stability of signal acquisition can be improved, providing a reliable data basis for subsequent analysis and identification.

[0050] When building a complete signal acquisition system, you also need to consider the system's data processing and storage methods, and choose appropriate data analysis software and storage devices to enable real-time monitoring, analysis, and recording of the collected signals.

[0051] By building a complete signal acquisition system, high-quality signal data can be obtained in real time, providing reliable technical support for engineering monitoring, fault diagnosis and predictive maintenance.

[0052] The signals collected in step 1 include temperature signals, pressure signals, vibration signals, electromagnetic signals, and ultrasonic signals; among them, the temperature signal is collected by infrared sensors, thermocouples, and thermal resistors installed on the GIS equipment housing, key contacts, and insulating material surfaces; the pressure signal is collected by piezoresistive pressure sensors and capacitive pressure sensors installed inside the GIS equipment to collect gas pressure and cooling system pressure; the vibration signal is collected by acceleration sensors and velocity sensors installed on the GIS equipment housing, connection parts, and supporting structures; the electromagnetic signal is collected by electromagnetic sensors, current transformers, and voltage transformers to collect current and voltage of GIS equipment; the ultrasonic signal is collected by ultrasonic sensors to detect wear and failure of GIS components.

[0053] Step 2, mapping PRPD map: generating PRPD map based on the collected signals;

[0054] Step 2 is implemented as follows:

[0055] PRPD (Partial Discharge Pattern) is a commonly used graph in partial discharge monitoring. By drawing a PRPD graph, the characteristics of partial discharge signals in GIS equipment can be intuitively displayed, which helps to identify and classify different types of partial discharges. The following are the detailed steps to generate a PRPD graph:

[0056] Step 2.1, signal preprocessing: Before generating the PRPD diagram, the collected signal data is preprocessed, including signal amplification, signal filtering, and signal denoising. In addition to preprocessing the collected signal data, it is also necessary to extract features and perform data analysis on the signal to obtain more comprehensive and accurate fault diagnosis results. By extracting and analyzing signal features, a PRPD diagram of foreign bodies inside the GIS is obtained. The PRPD diagram can intuitively display the characteristic patterns of different types of foreign bodies, which helps to quickly diagnose and determine the type of foreign body, thereby realizing the prediction and early warning of faults.

[0057] In summary, through signal preprocessing, feature extraction and data analysis, a more complete and accurate PRPD diagram can be constructed, which provides an important basis and reference for fault diagnosis and prediction.

[0058] Step 2.2, PRPD diagram generation: After signal preprocessing is completed, MATLAB software is used to generate the PRPD diagram. The PRPD diagram uses the peak amplitude and phase of the discharge pulse as coordinate axes, and each partial discharge pulse corresponds to a point in the diagram. By plotting the distribution of these points, the position characteristics of different partial discharge types in the PRPD diagram can be clearly displayed.

[0059] Step 2.3, data analysis: Analyze the generated PRPD map, and identify the partial discharge signals of different types of foreign objects inside the GIS by observing the aggregation and distribution patterns in different areas. Different discharge modes usually show unique distribution characteristics and aggregation on the PRPD map. Through observation and analysis, they can be accurately identified and distinguished. In addition to directly observing the discharge pattern characteristics on the PRPD map, other characteristic parameters such as pulse repetition frequency and energy distribution can also be combined for comprehensive analysis. By comprehensively considering these parameters, we can have a more comprehensive understanding of the nature and characteristics of partial discharge caused by foreign objects in the GIS, providing a more powerful basis for distinguishing foreign object types and maintenance decisions.

[0060] Step 2.4, identification and classification: The generated PRPD diagram can intuitively display the characteristics of local discharge signals in GIS equipment, providing an important reference for fault diagnosis and maintenance. For example, high-frequency discharge: It appears as a point with high amplitude and low phase in the diagram, which usually corresponds to the tip discharge or air gap discharge inside the GIS. Low-frequency discharge: It appears as a point with low amplitude and high phase in the diagram, which usually corresponds to the surface discharge of insulating materials. Periodic discharge: It appears as a regularly distributed point in the diagram, which usually corresponds to periodic defects in the equipment, such as wear of rotating machinery. Therefore, the generation of PRPD diagram is an important step and is of great significance for improving the safety and reliability of GIS equipment. The generation of PRPD diagram is to process the collected foreign body local discharge signal data and present them in a graphical way, which can help to make judgments and analyses quickly and accurately. In the PRPD diagram, the local discharge signals generated by different types of foreign bodies will show different characteristics, such as spikes and clusters. By observing these characteristics, different types of local discharges can be identified and classified, as follows:

[0061] In the PRPD graph, the characteristic points are concentrated in the narrow phase range (0°-30°) and (180°-210°), with large amplitude and large number of points. After identifying these points, it is confirmed to be bubble discharge;

[0062] In the PRPD diagram, the characteristic points are distributed in a wide phase range (30°-90°) and (210°-270°), with a small amplitude and a small number of points. After identifying these points, it is confirmed to be a dirty discharge;

[0063] In the PRPD graph, the characteristic points are distributed in the entire phase range, i.e., the narrow phase range (0°-360°), with small amplitude and moderate number of points. After identifying these points, it is confirmed to be a tip discharge;

[0064] In the PRPD diagram, the characteristic points are distributed in the entire phase range, that is, the wide phase range (0°-360°), with medium amplitude and moderate number of points. After identifying these points, it is confirmed to be suspended metal discharge.

[0065] In addition to using PRPD graphs to identify partial discharge signals, trend analysis of PRPD graphs is also performed to find changes in signal strength and frequency of occurrence, and to determine the working status of the equipment and the development trend of potential faults. Therefore, generating PRPD graphs not only helps to identify and classify the types of foreign matter in GIS equipment, but also plays an important role in equipment monitoring and maintenance, providing strong support for ensuring safe and reliable operation of equipment.

[0066] Step 3, grayscale recognition: use grayscale recognition algorithm to read the content in the PRPD image;

[0067] Step 3 is implemented as follows:

[0068] To use grayscale recognition to identify and extract features in PRPD images, the following steps are usually required:

[0069] Step 3.1, read PRPD image: Load PRPD image into MATLAB for further processing and analysis. Use MATLAB image processing to perform operations: including image enhancement, feature extraction, target detection, by processing PRPD image, extract useful information, understand the characteristics of partial discharge signal inside the equipment, in addition, use MATLAB for trend analysis can more intuitively display the data changes in PRPD image, help to find potential problems in time and take corresponding preventive measures.

[0070] Step 3.2, data preprocessing: Before grayscale recognition, the image needs to be preprocessed; for example, noise removal, contrast enhancement and other operations. Specifically include: 1. Noise removal, the purpose is to remove random noise in the image, retain effective signals, and improve image quality. The image can be smoothed by using a Gaussian filter to reduce high-frequency noise, the salt and pepper noise in the image can be removed by using a median filter, and the image can be smoothed while retaining the edge by using a bilateral filter; 2. Contrast enhancement, the purpose is to increase the contrast of the image and make the local discharge features more obvious; the grayscale value of the image can be evenly distributed by the histogram equalization method, the local image can be histogram equalized, the local contrast can be enhanced, the grayscale value range of the image can be linearly stretched, and the overall contrast can be improved; 3. Image sharpening, the purpose is to enhance the details in the image and make the local discharge features clearer. The Laplace operator can be used to enhance the edge of the image, and the difference with the original image after Gaussian blurring is performed, and the original image is superimposed to enhance the details; 4. Background removal, the purpose is to remove background information in the image and highlight the local discharge features. Static background is removed by background modeling and difference method. Use morphological operations such as opening and closing operations to remove small background noise; 5. Image cropping and scaling, the purpose is to crop and scale the image to meet the needs of subsequent processing and analysis. According to the location of the partial discharge area, the image is cropped, the key area is retained, and the image size is scaled as needed to meet the neural network input requirements; 6. Grayscale conversion, the purpose is to convert the color image into a grayscale image to simplify the subsequent processing steps. The color image can be converted into a grayscale image by using the weighted average method or the maximum value method; 7. Data standardization, the purpose is to standardize the image data and improve the model training effect. The image pixel value can be normalized to between 0-1 to increase the model convergence speed. Through the above detailed image preprocessing steps, the accuracy and reliability of partial discharge detection can be effectively improved, providing a high-quality data foundation for subsequent image analysis and processing.

[0071] Step 3.3, feature extraction: determine the features you want to extract from the PRPD image. For partial discharge recognition, the features include the number of groups, amplitude distribution, and phase distribution. Various methods can be used to extract these features from the image, such as shape detection.

[0072] Step 3.4, Grayscale Recognition: Once the required features are extracted, grayscale recognition algorithms (such as K-means clustering, support vector machines, etc.) can be used to classify and identify the features. The Statistics and Machine Learning Toolbox in MATLAB provides many functions and tools for implementing these algorithms.

[0073] Step 3.5: Result analysis: Analyze the classification results and evaluate the performance of the algorithm.

[0074] Step 4: CM-YOLOv3 algorithm classification: Identify the type of foreign matter inside the GIS cavity based on the read content.

[0075] Step 4 is implemented according to the following steps:

[0076] Foreign body type recognition is to detect foreign bodies inside the GIS cavity by applying the CM-YOLOv3 algorithm after generating the PRPD map. In order to meet the requirements of high-precision foreign body detection inside the GIS cavity and lightweight design of the network architecture, the present invention proposes the following Figure 2 The CM-YOLOv3 network structure shown.

[0077] The CM-YOLOv3 network structure includes the following:

[0078] 1. Input layer:

[0079] Input size: The image size is 416x416x3, which means the input is an RGB image with a size of 416x416 pixels; data preprocessing includes normalization and data enhancement;

[0080] 2. Convolutional layer and feature extraction layer:

[0081] A series of convolutional layers (Conv) and feature extraction layers are used in the CM-YOLOv3 network structure to extract the features of the input image. Specifically, they include:

[0082] Conv1: The convolution kernel size is 3x3, the number of output channels is 32, the stride is 1, and the padding is 1;

[0083] Batch Normalization (BN) and Leaky ReLU activation functions;

[0084] Conv2: The convolution kernel size is 3x3, the number of output channels is 64, the stride is 2, and the padding is 1;

[0085] Batch Normalization and Leaky ReLU;

[0086] Residual Block: Contains two convolutional layers (Conv3 and Conv4), each with a kernel size of 3x3, 64 output channels, a stride of 1, a padding of 1, and an addition connection method;

[0087] Conv5: The convolution kernel size is 3x3, the number of output channels is 128, the stride is 2, and the padding is 1;

[0088] BN and Leaky ReLU;

[0089] Residual Block: Contains two convolutional layers (Conv6 and Conv7), each with a kernel size of 3x3, 128 output channels, a stride of 1, a padding of 1, and an addition connection method;

[0090] Conv8: The convolution kernel size is 3x3, the number of output channels is 256, the stride is 2, and the padding is 1;

[0091] Batch Normalization and Leaky ReLU;

[0092] 3. Improvement module (CM module):

[0093] To improve detection accuracy and lightweight design, CM-YOLOv3 introduces an improved module (CM module), which includes:

[0094] Channel Attention Module (CAM): used to enhance the features of important channels;

[0095] AvgPool and MaxPool: perform global average pooling and maximum pooling on the input feature map and output feature vector;

[0096] FC1 and FC2: fully connected layers, input feature vectors, and output channel weights;

[0097] Sigmoid activation function: normalizes channel weights;

[0098] Channel-wise Multiplication: Multiply the channel weights by the original feature map channel by channel;

[0099] Spatial Attention Module (SAM): used to enhance the features of important spatial locations;

[0100] Concat: concatenate the average pooling and maximum pooling results of the input feature map in the channel dimension; Conv: the convolution kernel size is 7x7, the number of output channels is 1, the stride is 1, and the padding is 3;

[0101] Sigmoid activation function: normalizes spatial weights;

[0102] Element-wise Multiplication: Multiply the spatial weights by the original feature map element-wise;

[0103] 4. Detection Head:

[0104] CM-YOLOv3 uses three-scale detection heads to detect targets of different scales. The detection heads include:

[0105] Detection Head1: used to detect large-scale targets;

[0106] Conv9: The convolution kernel size is 1x1, the number of output channels is 256, the stride is 1, and the padding is 0;

[0107] Batch Normalization and Leaky ReLU;

[0108] Conv10: The convolution kernel size is 3x3, the number of output channels is 512, the stride is 1, and the padding is 1;

[0109] Batch Normalization and Leaky ReLU;

[0110] Conv11: The convolution kernel size is 1x1, the number of output channels is 255, the stride is 1, and the padding is 0;

[0111] Output: feature map size is 13x13x255;

[0112] Detection Head2: used to detect mesoscale targets;

[0113] Conv12: The convolution kernel size is 1x1, the number of output channels is 128, the stride is 1, and the padding is 0;

[0114] Batch Normalization and Leaky ReLU;

[0115] Conv13: The convolution kernel size is 3x3, the number of output channels is 256, the stride is 1, and the padding is 1;

[0116] Batch Normalization and Leaky ReLU;

[0117] Conv14: The convolution kernel size is 1x1, the number of output channels is 255, the stride is 1, and the padding is 0;

[0118] Output: feature map size is 26x26x255;

[0119] Detection Head3: used to detect small-scale targets;

[0120] Conv15: The convolution kernel size is 1x1, the number of output channels is 64, the stride is 1, and the padding is 0;

[0121] BN and Leaky ReLU;

[0122] Conv16: The convolution kernel size is 3x3, the number of output channels is 128, the stride is 1, and the padding is 1;

[0123] Batch Normalization and Leaky ReLU;

[0124] Conv17: convolution kernel size is 1x1, number of output channels is 255, stride is 1, padding is 0; output: feature map size is 52x52x255;

[0125] 5. Output layer:

[0126] Bounding Box and Class Scores: Outputs the bounding box and category confidence generated by each detection head;

[0127] Non-Maximum Suppression (NMS): Perform NMS processing on the test results to remove duplicate detections;

[0128] The NMS processing in step 4 is as follows:

[0129] Firstly, the K-means algorithm is used to cluster the anchor frames of foreign objects to obtain the anchor frame values ​​of the foreign objects to be detected, so that the algorithm can regress the bounding box according to the specific size of the foreign objects; drawing on the idea of ​​CSPNet network, CSP-Darknet36 is designed as a new backbone network to increase the gradient path in the backbone network, reduce the repeated gradient information in the network, and improve the network learning ability;

[0130] The specific structure of CSP-Darknet36 is:

[0131] CSP Block1:

[0132] Conv5: kernel size is 3x3, number of output channels is 128, stride is 2, padding is 1. BN and Leaky ReLU.

[0133] CSP branch 1: It is divided into two parts, the trunk branch and the residual branch.

[0134] Main branch: Contains two convolutional layers (Conv6 and Conv7), each with a kernel size of 3x3, 128 output channels, a stride of 1, a padding of 1, and an addition connection method.

[0135] Residual branch: directly connects the input and output.

[0136] Concatenation: Concatenate the trunk branch and the residual branch in the channel dimension.

[0137] Conv8: The convolution kernel size is 1x1, the number of output channels is 128, the stride is 1, and the padding is 0.

[0138] CSP Block2:

[0139] Conv9: kernel size is 3x3, number of output channels is 256, stride is 2, padding is 1. BN and Leaky ReLU.

[0140] CSP branch 2: It is divided into two parts, the trunk branch and the residual branch.

[0141] Main branch: Contains two convolutional layers (Conv10 and Conv11), each with a kernel size of 3x3, 256 output channels, a stride of 1, a padding of 1, and an addition connection method.

[0142] Residual branch: directly connects the input and output.

[0143] Concatenation: Concatenate the trunk branch and the residual branch in the channel dimension.

[0144] Conv12: The convolution kernel size is 1x1, the number of output channels is 256, the stride is 1, and the padding is 0.

[0145] CSP Block3:

[0146] Conv13: kernel size is 3x3, number of output channels is 512, stride is 2, padding is 1. BN and LeakyReLU.

[0147] CSP branch 3: It is divided into two parts, the trunk branch and the residual branch.

[0148] Main branch: Contains two convolutional layers (Conv14 and Conv15), each with a kernel size of 3x3, 512 output channels, a stride of 1, a padding of 1, and an addition connection method.

[0149] Residual branch: directly connects the input and output.

[0150] Concatenation: Concatenate the trunk branch and the residual branch in the channel dimension.

[0151] Conv16: The convolution kernel size is 1x1, the number of output channels is 512, the stride is 1, and the padding is 0.

[0152] CSP Block4:

[0153] Conv17: kernel size is 3x3, number of output channels is 1024, stride is 2, padding is 1. BN and LeakyReLU.

[0154] CSP branch 4: It is divided into two parts, the trunk branch and the residual branch.

[0155] Main branch: Contains two convolutional layers (Conv18 and Conv19), each with a kernel size of 3x3, 1024 output channels, a stride of 1, a padding of 1, and an addition connection method.

[0156] Residual branch: directly connects the input and output.

[0157] Concatenation: Concatenate the trunk branch and the residual branch in the channel dimension.

[0158] Conv20: The convolution kernel size is 1x1, the number of output channels is 1024, the stride is 1, and the padding is 0.

[0159] Output layer:

[0160] Output feature map: Feature maps are extracted from each output layer of the CSP module for object detection by the subsequent detection head.

[0161] Combination Figure 3 , Figure 4 , Figure 5 ,Multi-scale feature fusion: fuse feature maps of different scales to generate the final feature map.

[0162] A new MFPN feature pyramid is proposed, which enables the algorithm to utilize more position information in the shallow network and semantic information in the deep network;

[0163] The specific structure of the MFPN feature pyramid is:

[0164] 1. Input layer

[0165] Input feature maps: Output feature maps of different scales from the CSP-Darknet36 backbone network.

[0166] Input Features Figure 1 (P1): size is 13x13x1024.

[0167] Input Features Figure 2 (P2): Dimensions are 26x26x512.

[0168] Input Features Figure 3 (P3): size is 52x52x256.

[0169] 2. Upsampling and feature fusion layer

[0170] MFPN achieves multi-scale fusion and enhancement of features through upsampling and feature fusion layers.

[0171] UpSampling layer: Upsample the low-scale feature map to make its size match the high-scale feature map.

[0172] UpSample1: Upsample P1 to an output size of 26x26x1024.

[0173] UpSample2: Upsample P2 to an output size of 52x52x512.

[0174] Feature Fusion Layer: The upsampled feature map is fused with the high-scale feature map by using the element-wise addition and concatenation methods.

[0175] Fusion1: Fusion UpSample1 with P2, output size is 26x26x1536 (1024+512). Conv21: Convolution kernel size is 1x1, output channels are 512, stride is 1, padding is 0. BN and Leaky ReLU.

[0176] Fusion2: Fusion UpSample2 with P3, output size is 52x52x768 (512+256). Conv22: Convolution kernel size is 1x1, output channels are 256, stride is 1, padding is 0. BN and Leaky ReLU.

[0177] 3. Feature Enhancement Layer

[0178] Further feature enhancement is performed on the fused feature map to improve the feature expression capability.

[0179] Enhancement layer 1 (Enhance1): Convolution operation is performed on Fusion1 to further extract features.

[0180] Conv23: kernel size is 3x3, number of output channels is 512, stride is 1, padding is 1. BN and LeakyReLU.

[0181] Conv24: kernel size is 3x3, number of output channels is 512, stride is 1, padding is 1. BN and LeakyReLU.

[0182] Enhancement layer 2 (Enhance2): Convolution operation is performed on Fusion2 to further extract features.

[0183] Conv25: kernel size is 3x3, number of output channels is 256, stride is 1, padding is 1. BN and LeakyReLU.

[0184] Conv26: kernel size is 3x3, number of output channels is 256, stride is 1, padding is 1. BN and LeakyReLU.

[0185] 4. Downsampling and feature integration layer

[0186] Through downsampling and feature integration layers, multi-scale integration and enhancement of features are achieved.

[0187] DownSampling layer: Downsample the high-scale feature map to make its size match the low-scale feature map.

[0188] DownSample1: Downsample P3 to an output size of 26x26x256.

[0189] Conv27: kernel size is 3x3, number of output channels is 256, stride is 2, padding is 1. BN and LeakyReLU.

[0190] DownSample2: Downsample P2 to an output size of 13x13x512.

[0191] Conv28: kernel size is 3x3, number of output channels is 512, stride is 2, padding is 1. BN and LeakyReLU.

[0192] Feature Integration Layer: Integrate the downsampled feature map with the low-scale feature map by element-by-element addition and concatenation.

[0193] Integration1: Integrate DownSample1 and Fusion1, and the output size is 26x26x768 (256+512).

[0194] Conv29: kernel size is 1x1, number of output channels is 512, stride is 1, padding is 0. BN and LeakyReLU.

[0195] Integration2: Integrate DownSample2 and Fusion2, and the output size is 13x13x1024 (512+512).

[0196] Conv30: kernel size is 1x1, number of output channels is 1024, stride is 1, padding is 0. BN and LeakyReLU.

[0197] 5. Output Layer

[0198] Multi-scale feature map: Feature maps are extracted from each output layer of MFPN for object detection in subsequent detection heads.

[0199] Output characteristics Figure 1 : The size is 13x13x1024.

[0200] Output characteristics Figure 2 : The dimensions are 26x26x512.

[0201] Output characteristics Figure 3 : The size is 52x52x256.

[0202] Step 4.1 Foreign body size clustering:

[0203] The YOLOv3 algorithm uses the anchor frame mechanism proposed by Faster R-CNN. The number and size of the anchor frame affect the accuracy of target detection. During the training process, the parameters of the candidate frame are continuously adjusted to make it closer to the true value frame. The YOLOv3 algorithm uses K-means clustering to obtain the optimal number and size of anchor frames for the COC0 dataset. The characteristics of foreign objects inside GIS equipment are very different from the target characteristics of the COC0 dataset. In order to make the convolutional neural network better learn the characteristics of foreign objects, it is necessary to cluster the training samples to obtain the number and size of anchor frames suitable for foreign objects inside the GIS cavity.

[0204] The traditional K-means algorithm uses Euclidean distance as the objective function, which will produce a "large box advantage", that is, a large bounding box will produce a larger error than a small bounding box. However, IOU (intersection over union) has nothing to do with the size of the bounding box. Therefore, using IOU to participate in the distance formula calculation can avoid this phenomenon. The experiment uses a new distance formula (1) instead of Euclidean clustering. Box represents the clustering result, and centroid represents the cluster center:

[0205] d(box,centroid)=1-10U(bpx,centroid) (1)

[0206] By clustering the foreign body samples, the relationship curve between the average intersection-over-union ratio of foreign body samples and the number of anchor frames is as follows: Figure 3 shown.

[0207] It can be seen from the figure that as the number of anchor boxes increases, the average intersection-union ratio gradually increases. When the number of anchor boxes is greater than 9, the average intersection-union ratio increases slowly. Therefore, the number of anchor boxes is selected as 9, and K-means clustering is used to obtain the best nine groups of anchor box values, which are: (10.32), (13.14), (18.22), (19.41), (22, 35), (35, 59), (56, 74), (65, 98), (82, 54);

[0208] Step 4.2, Improved backbone network CSP-Darknet36:

[0209] YOLOv3 uses Darknet-53 as the feature extraction network and utilizes the residual structure to extract the feature information of the target. However, due to the relatively complex network structure, the overall network parameters are too large, which is not ideal in terms of practicality of foreign body detection. Therefore, the present invention chooses to design a new backbone network to improve the learning rate of the algorithm and reduce the learning parameters.

[0210] CSPNet is a network used to enhance the performance of convolutional neural networks. It can be widely used in convolutional neural networks such as ResNet, ResNeXt and DenseNet. The CSPNet network structure is as follows: Figure 4 As shown, Figure 4 (a) is the original convolutional neural network. Figure 4 (b) is the CSPNet network. CSPNet divides the propagation path of the base layer features into two parts, reduces the learning parameters by reducing the number of convolutional layer channels, and increases the learning ability by reducing the gradient reuse information.

[0211] The specific network structure of CSPNet is as follows:

[0212] 1. Input layer

[0213] Input image: size is HxWx3 (e.g., 416x416x3).

[0214] 2. Backbone network structure

[0215] CSPNet uses CSP-Darknet36 as the backbone network, including multiple convolutional layers, residual blocks, and cross-stage partial connections. The following is the specific hierarchy and connection method:

[0216] Phase 1:

[0217] Conv1: The convolution kernel size is 3x3, the number of output channels is 32, the stride is 1, and the padding is 1.

[0218] Batch Normalization and Leaky ReLU.

[0219] Conv2: The convolution kernel size is 3x3, the number of output channels is 64, the stride is 2, and the padding is 1.

[0220] Batch Normalization and Leaky ReLU.

[0221] Phase 2:

[0222] Residual Block 1: Contains two convolutional layers, each with a convolution kernel size of 3x3, 64 output channels, a stride of 1, and a padding of 1.

[0223] Conv3 and Conv4, each layer is followed by BN and Leaky ReLU.

[0224] CSP Connection 1: Cross-stage partial connection, concatenating the output of residual block 1 and the output of Conv2, and then performing feature fusion through a convolutional layer. Conv5: The convolution kernel size is 1x1, the number of output channels is 128, the stride is 1, and the padding is 0. BN and Leaky ReLU.

[0225] Phase 3:

[0226] Conv6: kernel size is 3x3, number of output channels is 128, stride is 2, padding is 1. BN and Leaky ReLU.

[0227] Residual Block 2: Contains two convolutional layers, each with a convolution kernel size of 3x3, 128 output channels, a stride of 1, and a padding of 1.

[0228] Conv7 and Conv8, each layer is followed by BN and Leaky ReLU.

[0229] CSP Connection 2: A cross-stage partial connection, concatenating the output of residual block 2 and the output of Conv6, and then performing feature fusion through a convolutional layer.

[0230] Conv9: kernel size is 1x1, number of output channels is 256, stride is 1, padding is 0. BN and Leaky ReLU.

[0231] Phase 4:

[0232] Conv10: kernel size is 3x3, number of output channels is 256, stride is 2, padding is 1. BN and LeakyReLU.

[0233] Residual Block 3: Contains two convolutional layers, each with a convolution kernel size of 3x3, 256 output channels, a stride of 1, and a padding of 1.

[0234] Conv11 and Conv12, each layer is followed by BN and Leaky ReLU.

[0235] CSP Connection 3: A cross-stage partial connection, concatenating the output of residual block 3 and the output of Conv10, and then performing feature fusion through a convolutional layer.

[0236] Conv13: kernel size is 1x1, number of output channels is 512, stride is 1, padding is 0. BN and LeakyReLU.

[0237] Phase 5:

[0238] Conv14: kernel size is 3x3, number of output channels is 512, stride is 2, padding is 1. BN and LeakyReLU.

[0239] Residual Block 4: Contains two convolutional layers, each with a convolution kernel size of 3x3, 512 output channels, a stride of 1, and a padding of 1.

[0240] Conv15 and Conv16, each layer is followed by BN and Leaky ReLU.

[0241] CSP Connection 4: A cross-stage partial connection that concatenates the output of residual block 4 and the output of Conv14, and then performs feature fusion through a convolutional layer.

[0242] Conv17: kernel size is 1x1, number of output channels is 1024, stride is 1, padding is 0. BN and LeakyReLU.

[0243] 3. Output layer

[0244] Global Average Pooling: Perform global average pooling on the final feature map and output a fixed-length vector.

[0245] Fully Connected Layer: The result of global average pooling is input into the fully connected layer and the classification result is output.

[0246] FC1: Fully connected layer, the number of output channels is 1000 (adjustable according to specific tasks).

[0247] Softmax activation function.

[0248] Considering the high real-time requirements of foreign body detection in GIS cavity and the limited computing power of industrial computers, the learning parameters of the network should be further reduced. The present invention draws on the CSPNet network structure to design a new backbone network. The CSPNet structure is applied to the original backbone network Darknet-53, and the network model parameters are reduced by reducing the number of convolutional layer channels; the gradient propagation path is increased, and the gradient splitting and merging strategy is used to reduce the reuse of gradient information, thereby increasing the learning ability of the convolutional neural network;

[0249] In order to further improve the detection speed, the present invention decided to simplify the feature extraction network model. Appropriate deletion of some convolutional layers can improve the detection speed of the model. The Darknet53 network has a total of 5 residual blocks, each of which has 1, 2, 88, and 4 residual units respectively. The relatively redundant residual units of the 3rd, 4th, and 5th residual blocks are deleted to 4, 4, and 2 to form an improved feature extraction network, called CSP-Darknet36, whose structure diagram is shown in the figure. Figure 6 As shown;

[0250] Step 4.3, improved feature pyramid MFPN:

[0251] Since the light inside the GIS cavity is dim and the size of foreign objects is small, the feature information contained in the foreign objects is relatively small. Although YOLOv3 uses the FPN feature pyramid to fuse the multi-scale features of the target, the reduction in feature information causes the features learned by the algorithm to be not broad enough, resulting in poor detection effect.

[0252] The FPN feature pyramid ultimately only uses the semantic information in the deep network, and the position information in the shallow network is not used. Acquiring the position information in the shallow network can improve the algorithm detection performance. Therefore, the present invention chooses to improve the original feature pyramid, using the three-scale feature information fusion method, first downsampling the feature maps of 104*104, 52*52, and 26*26 scales, and then merging them with the 52*52, 26*26, and 13*13 feature maps before the output layer, and then upsampling with the 26*26 and 13*13 feature maps for sub-combination, and finally outputting feature information of three scales.

[0253] Compared with the original feature pyramid, the improved feature pyramid integrates more levels of feature information, not only using the position information in the shallow network such as 104*104, but also learning the semantic feature information in the deep network such as 13*13, which better improves the foreign body detection effect. The present invention calls the improved feature pyramid MFPN, and the structural comparison diagram of MFPN and FPN is shown in the figure.

[0254] Example 1

[0255] The present invention compares four models: YOLOv3, CSP36-YOLOv3, MFPN-YOLOv3, and CM-YOL0v3. CSP36-YOLOV3 is the YOLOv3 algorithm after modifying the backbone network; MFPN-YOLOv3 is the YOLOv3 algorithm after modifying the feature pyramid.

[0256] In deep learning, mAP (mean average precision) is generally used as an indicator to measure the detection accuracy in target detection. First, AP (average precision) is calculated by P (precision) R (recall), and then the AP values ​​of all categories are averaged to get mAP. The precision and recall calculation formulas are shown in formulas (2) and (3); TP represents the number of targets correctly predicted as positive examples when IOU (the ratio of the intersection area of ​​the predicted target box and the true value box to the union area) is greater than 0.5; FP represents the number of targets that are incorrectly predicted as positive examples and whose IOU is less than 0.5; FN represents the number of targets that are not detected but are actually positive examples.

[0257]

[0258] The PR curve is calculated by precision and recall, and then the AP value is obtained by integrating the PR curve. The mean average precision (mAP) is obtained by mAP calculation formula (4). Where OR represents the total number of detection categories.

[0259]

[0260] Example 2

[0261] Detection performance comparison

[0262] Four algorithm models are used to detect foreign objects in the test set. The precision and recall rates are shown in Table 1. It can be seen from Table 1 that the improved CM-YOLOv3 algorithm has obvious improvements in recall rate, precision rate and F1 score.

[0263] Table 1 Comparison of precision and recall

[0264] Network Name Accuracy (%) Recall rate (%) F1 score YOLOv3 88 80 84 CSP36-YOLOv3 91 83 87 MFPN-YOLOv3 92 83 87 CM-YOLOv3 93 85 89

[0265] The average accuracy and mean average precision values ​​of different foreign objects under the four algorithm models are shown in Table 2.

[0266] Table 2 Comparison of foreign AP and mAP in different networks

[0267]

[0268] As can be seen from Table 2, the improved CM-YOLOv3 algorithm has improved the average accuracy of each foreign body detection. The mAP values ​​of CSP36-YOLOv3, MFPN-YOLOv3, and CM-YOLOv3 algorithms for foreign body detection in GIS cavities are 82.12%, 80.80%, and 84.32%, respectively, which are 3.97%, 2.65%, and 6.17% higher than the original YOLOv3 algorithm, respectively. This shows that the CSP-Darknet36 backbone network and MFPN feature pyramid module proposed in the present invention are effective in improving the performance of foreign body detection. And the improved CM-YOLOv3 algorithm has the most obvious improvement in the detection accuracy of screw foreign bodies. Among the four foreign bodies, screws have more types and sizes, are relatively smaller in size, and have less feature information. The improved CM-YOLOv3 algorithm enhances the learning ability of the convolutional neural network by introducing the CSPNet network; improves the feature pyramid, integrates the features and location information of targets at multiple levels; and improves the detection ability of small target foreign bodies.

[0269] Example 3

[0270] Comparison of detection efficiency

[0271] The network weight result file stores the weight parameters of each network layer during the training process, and its size represents the overall complexity of the network structure. The sizes of the weight model files trained by the four network types are shown in Table 3.

[0272] Table 3 Network weight comparison table

[0273] Network Name Weight size (MB) YOLOv3 240.59 CSP36-YOLOv3 148.56 MFPN-YOLOv3 241.61 CM-YOLOv3 151.89

[0274] From the table, we can see that the weight file size of the CM-YOLOv3 network is reduced by 37% compared to the YOLOv3 network. The reduction in the overall computational workload of the network can reduce the algorithm overhead.

[0275] When implementing this method, according to the specific monitoring requirements and scene characteristics, suitable sensors are selected for signal acquisition, such as voltage sensors, current sensors, pressure sensors, etc. The signals collected by these sensors will be transmitted to the computer for subsequent processing. In the computer, by processing the collected signals, a PRPD diagram can be generated, which is a two-dimensional image, with the horizontal axis representing the pulse peak charge or amplitude and the vertical axis representing the pulse peak time difference or phase.

[0276] After generating the PRPD spectrum, the grayscale recognition algorithm can be used to identify the features in the spectrum. By extracting features in the spectrum, such as amplitude, frequency distribution, etc., partial discharge can be better identified and analyzed. Finally, the specific type of partial discharge is identified by the support vector machine classification algorithm. Support vector machine is an effective classification algorithm that can accurately classify different types of partial discharge signals, providing important support for preventing equipment failures and improving system reliability.

[0277] In summary, when implementing this method, by selecting suitable sensors, generating PRPD images, using grayscale recognition algorithms to identify features, and using the CM-YOLOv3 algorithm to detect foreign objects inside the GIS cavity, reliable guarantees are provided for the safe operation of the equipment.

[0278] This method can accurately identify different types of partial discharges caused by different foreign objects by subsequent processing of the collected signals, generating PRPD images and using grayscale recognition algorithms for feature recognition. In this way, the monitoring effect of foreign objects inside the GIS cavity can be effectively improved, and potential faults can be discovered in a timely manner. The reliability and accuracy of this method make it widely used in the industrial field and play an important role in equipment management and maintenance.

[0279] Aiming at the complex and low detection accuracy of foreign body detection methods in GIS equipment, this paper proposes a target detection algorithm based on CM-YOLOv3. Through experiments and analysis, it is concluded that the type of foreign body can be accurately identified, so this method is highly efficient and widely applicable, and has the potential for large-scale promotion.

Claims

1. A GIS internal foreign body defect monitoring method based on improved YOLOv3 is characterized by: Follow the steps below to implement it: Step 1: Collect signals: collect signals from high-voltage equipment through sensors; Step 2, mapping PRPD map: generating PRPD map based on the collected signals; Step 3, grayscale recognition: use grayscale recognition algorithm to read the content in the PRPD image; Step 4: CM-YOLOv3 algorithm classification: Identify the type of foreign matter inside the GIS cavity based on the read content.

2. The GIS internal foreign body defect monitoring method based on improved YOLOv3 according to claim 1 is characterized in that: The step 1 is specifically as follows: The signals collected in step 1 include temperature signals, pressure signals, vibration signals, electromagnetic signals, and ultrasonic signals; among them, the temperature signal is collected by infrared sensors, thermocouples, and thermal resistors installed on the GIS equipment housing, key contacts, and insulating material surfaces; the pressure signal is collected by piezoresistive pressure sensors and capacitive pressure sensors installed inside the GIS equipment to collect gas pressure and cooling system pressure; the vibration signal is collected by acceleration sensors and velocity sensors installed on the GIS equipment housing, connection parts, and supporting structures; the electromagnetic signal is collected by electromagnetic sensors, current transformers, and voltage transformers to collect current and voltage of GIS equipment; the ultrasonic signal is collected by ultrasonic sensors to detect wear and failure of GIS components.

3. The GIS internal foreign body defect monitoring method based on improved YOLOv3 according to claim 2 is characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, signal preprocessing: Before generating the PRPD graph, the collected signal data is preprocessed, including signal amplification, signal filtering, and signal denoising. In addition to preprocessing the collected signal data, it is also necessary to extract features and perform data analysis on the signal; Step 2.2, PRPD diagram generation: After the signal preprocessing is completed, the PRPD diagram is generated using MATLAB software. The PRPD diagram uses the peak amplitude and phase of the discharge pulse as coordinate axes, and each partial discharge pulse corresponds to a point in the diagram; Step 2.3, data analysis: Analyze the generated PRPD map, and identify the partial discharge signals of different types of foreign objects inside the GIS by observing the aggregation and distribution patterns in different areas; Step 2.4, identification and classification: The generated PRPD diagram intuitively displays the characteristics of partial discharge signals in GIS equipment, providing an important reference for fault diagnosis and maintenance work; in the PRPD diagram, the partial discharge signals generated by different types of foreign matter will show different characteristics. By observing these characteristics, different types of partial discharges can be identified and classified, as follows: In the PRPD image, the characteristic points are concentrated in the narrow phase range (0°-30°) and (180°-210°). After identifying these points, it is confirmed to be bubble discharge; In the PRPD diagram, the characteristic points are distributed in a wide phase range (30°-90°) and (210°-270°), and after identifying these points, it is confirmed to be a contaminated discharge; In the PRPD graph, the characteristic points are distributed in the entire phase range, i.e., the narrow phase range (0°-360°), and after identifying these points, it is confirmed to be a tip discharge; In the PRPD diagram, the characteristic points are distributed in the entire phase range, i.e., the wide phase range (0°-360°). After identifying these points, it is confirmed that it is suspended metal discharge; By performing trend analysis on the PRPD graph, changes in signal strength and frequency of occurrence can be discovered, and the working status of the equipment and the development trend of potential faults can be determined.

4. The GIS internal foreign body defect monitoring method based on improved YOLOv3 according to claim 3 is characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1, read PRPD image: load PRPD image into MATLAB, and use MATLAB image processing to perform operations: including image enhancement, feature extraction, and target detection. By processing the PRPD image, useful information can be extracted to understand the characteristics of the partial discharge signal inside the equipment. In addition, trend analysis using MATLAB can intuitively display the data changes in the PRPD image, help to timely discover potential problems and take corresponding preventive measures; Step 3.2, data preprocessing: Before grayscale recognition, the image needs to be preprocessed; Step 3.3, feature extraction: determine the features you want to extract from the PRPD image. For partial discharge identification, the features include the number of groups, amplitude distribution, and phase distribution. Step 3.4, grayscale recognition: Use grayscale recognition algorithm to classify and identify features; Step 3.5: Result analysis: Analyze the classification results and evaluate the performance of the algorithm.

5. The GIS internal foreign body defect monitoring method based on improved YOLOv3 according to claim 4 is characterized in that: The specific details of the foreign body size clustering in step 4.1 are as follows: The new distance formula (1) is used to replace the Euclidean clustering. Box represents the clustering result, and centroid represents the cluster center: d(box,centroid)=1-10U(bpx,centroid) (1) By clustering analysis of foreign body samples, it is found that the average intersection-union ratio gradually increases with the increase of the number of anchor frames. When the number of anchor frames is greater than 9, the average intersection-union ratio increases slowly. Therefore, the number of anchor frames is selected as 9, and K-means clustering is used to obtain the best nine groups of anchor frame values, which are: (10.32), (13.14), (18.22), (19.41), (22, 35), (35, 59), (56, 74), (65, 98), (82, 54); Step 4.2, Improved backbone network CSP-Darknet36: The CSPNet structure is applied to the original backbone network Darknet-53, and the network model parameters are reduced by reducing the number of convolutional layer channels; the gradient propagation path is increased and the gradient splitting and merging strategy is used to reduce the reuse of gradient information and increase the learning ability of the convolutional neural network; The Darknet53 network has a total of 5 residual blocks, each of which has 1, 2, 88, and 4 residual units respectively. The residual units of the relatively redundant 3rd, 4th, and 5th residual blocks are deleted to 4, 4, and 2 to form an improved feature extraction network called CSP-Darknet36; Step 4.3, improved feature pyramid MFPN: Improvements are made on the original feature pyramid by fusing three-scale feature information. The feature maps of 104*104, 52*52, and 26*26 scales are first downsampled, and then fused with the 52*52, 26*26, and 13*13 feature maps before the output layer. Finally, they are upsampled with the 26*26 and 13*13 feature maps and combined to output feature information of three scales.

Images