Distribution cabinet equipment inspection system based on residual network

By combining the residual network of deep and shallow network models, the problem of inaccurate identification caused by incomplete feature extraction and training complexity in power distribution cabinet equipment inspection is solved, and the rapid and accurate status judgment of power distribution cabinet equipment components is achieved. and early warning improves the safety and efficiency of the equipment.

CN118823289BActive Publication Date: 2024-11-22ZHEJIANG JINDUN TECH CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411302889.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-19
Publication Date
2024-11-22
Estimated Expiration
2044-09-19

AI Technical Summary

Technical Problem

In the inspection of traditional distribution cabinet equipment, it is difficult for shallow network models to collect complex features, resulting in large detection errors, while the complex training of deep network models leads to an increase in network depth, resulting in inaccurate identification of component status of distribution cabinet equipment and inability to provide timely feedback and early warnings. There are safety hazards.

Method used

The inspection system based on residual network is adopted, combined with deep network and shallow network models for feature extraction, and through residual block fusion features, a complete residual network model is trained to ensure the data accuracy when the network depth complexity increases, and to achieve Accurate identification and feedback on the component status of the distribution cabinet equipment.

Benefits of technology

It improves the completeness and accuracy of feature extraction, quickly distinguishes the abnormal state of the distribution cabinet equipment components, realizes more accurate and intelligent state judgment, reduces detection errors, and improves the safety and efficiency of the equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118823289B_ABST
    Figure CN118823289B_ABST
Patent Text Reader

Abstract

The present invention discloses a distribution cabinet equipment inspection system based on a residual network, which relates to the technical field of distribution cabinet detection, and includes a data acquisition and preprocessing unit, a depth analysis unit, a state recognition unit and an inspection feedback unit; the technical key points are: preprocessing the collected distribution cabinet equipment image, and extracting the first feature result of the preprocessed image data based on a deep network model, and extracting the second feature result of the preprocessed image data based on a shallow network model to train the residual network model, so as to ensure that the collected features are accurate enough and reduce the detection error; the residual network model involves a convolutional layer, a fully connected layer, a pooling layer and a residual block, which can effectively solve the complexity of the network depth, improve the training efficiency, and analyze the target state of the distribution cabinet equipment components more quickly to distinguish the abnormal state for subsequent monitoring and early warning work.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution cabinet detection, and in particular to a power distribution cabinet equipment inspection system based on a residual network. Background Art

[0002] In the inspection of the distribution room, the traditional method relies on periodic physical inspection to identify the components of the distribution cabinet equipment. With the development of technology, intelligent inspection systems such as "intelligent inspection robots" have emerged; especially in the distribution cabinet equipment, the performance and status of each distribution cabinet equipment component have a direct impact on the stability and efficiency of the entire equipment; in terms of feature recognition of images, one is a shallow network feature extraction model, such as: image recognition based on artificially designed features, involving fewer layers; the other is a deep network feature extraction model, such as: image recognition based on deep learning automatic feature extraction method, involving more layers;

[0003] However, when inspecting traditional power distribution cabinet equipment:

[0004] Image recognition often relies on shallow network models to extract features. It is difficult to collect some complex features, and in the subsequent process, the model is trained based on the collected features, resulting in a lack of accuracy in the entire process and easily causing large detection errors. The process of extracting models based on deep network features often increases the depth of the network due to the complexity of the training process, resulting in training difficulties. This ultimately leads to inaccurate identification of the status of distribution cabinet equipment components. When the distribution cabinet equipment components are in an abnormal state, it is impossible to provide timely feedback and warnings, which can easily lead to safety hazards. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention provides a distribution cabinet equipment inspection system based on a residual network. By combining two different methods: one is a method of feature extraction based on a deep network model, and the other is a method of feature extraction based on a shallow network model, the integrity of feature collection in the system training process is effectively solved. At the same time, a residual block is designed to fuse the features to train a complete residual network model, and the accuracy of the data is ensured as the network depth and complexity increase. Then, the status of the distribution cabinet equipment components is identified to implement different feedback responses, thereby solving the problems in the background technology.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A distribution cabinet equipment inspection system based on a residual network includes: a data acquisition and preprocessing unit, a deep analysis unit, a state identification unit and an inspection feedback unit, and the data acquisition and preprocessing unit, the deep analysis unit, the state identification unit and the inspection feedback unit are communicatively connected with each other;

[0010] The data acquisition and preprocessing unit uses image acquisition technology to collect image data of the power distribution cabinet equipment in real time and preprocess the image data;

[0011] A depth analysis unit is configured to preset a residual network model, wherein a deep network model and a shallow network model are built in the residual network model; and based on the deep network model, feature extraction is performed on the preprocessed image data to obtain a first feature result; and at the same time, based on the shallow network model, feature extraction is performed on the preprocessed image data to obtain a second feature result; and the residual network model is trained according to the first feature result and the second feature result to obtain an updated residual network model for the next feature extraction;

[0012] The state recognition unit is based on the residual network model. It extracts features through several convolutional layers, and fuses the features extracted by the shallow network model with the features extracted by the deep network model through the residual block. Then, the feature dimension is processed through the pooling layer. Finally, the fully connected layer is used to map the extracted features to the category space, and the output layer is used to output the predicted value.

[0013] The inspection feedback unit determines the target state of the distribution cabinet equipment components based on the predicted value. The target state includes a normal state or an abnormal state. When it is determined to be a normal state, no response feedback action is taken; when it is determined to be an abnormal state, a response feedback action is taken.

[0014] Furthermore, the process of preprocessing the image data is as follows:

[0015] Image recognition: Preset and build a device component library, analyze the collected image data, crop the target area of ​​the device component, and generate the device component image; the device component image is in RGB color format;

[0016] Color model: Grayscale processing: Use the average method to convert the device component image into a grayscale image; Morphological operation: Use the opening operation based on the grayscale image to remove shadows; Binarization segmentation: After the opening operation, perform binary classification so that its grayscale value is 0 or 1.

[0017] Furthermore, the process of obtaining the first feature result is as follows:

[0018] Feature extraction: extract features from the preprocessed image to form feature vectors, and concatenate the feature vectors to form a state feature matrix; the feature vectors include but are not limited to color features and texture features;

[0019] Matrix analysis: pre-set the feature matrix threshold, compare and analyze the state feature matrix corresponding to each device component image with the preset feature matrix threshold, and when the state feature matrix meets the preset feature matrix threshold, the corresponding position is recorded as 1, otherwise it is recorded as 0; and mark them in turn to form the first matrix r1;

[0020] Output result: Integrate the first matrix r1 and mark it as the first feature result.

[0021] Furthermore, based on the shallow network model, feature extraction is performed on the image data to form a feature vector, and a second matrix r2 is constructed, and the second matrix r2 is integrated and marked as a second feature result.

[0022] Further, the first feature result includes a first data set U1 obtained according to the deep network model that meets the preset standard, a second data set U2 that does not meet the preset standard, and a third data set U3 that cannot be judged whether it meets the preset standard; the second feature result includes a fourth data set U4 obtained according to the shallow network model that meets the preset standard and a fifth data set U5 that does not meet the preset standard; then the process of training the residual network model based on the first feature result and the second feature result is:

[0023] Marking identification: assigning corresponding identifications to the first data set U1, the second data set U2, the third data set U3, the fourth data set U4 and the fifth data set U5; wherein, the data that meet the preset standard are marked as the first identification g1, the data that do not meet the preset standard are marked as the second identification g2, and the data that cannot be judged whether it meets the preset standard are marked as the third identification g3;

[0024] Determination analysis:

[0025] First judgment: perform judgment analysis on the first data set U1 and the fourth data set U4 to obtain the corresponding sixth data set U6 and seventh data set U7; and use the sixth data set U6, the seventh data set U7 and the first identifier g1 as input information to train the residual network model;

[0026] The second data set U2 and the fifth data set U5 are subjected to judgment analysis to obtain the corresponding eighth data set U8 and ninth data set U9; and the eighth data set U8, the ninth data set U9 and the second identifier g2 are used as input information to train the residual network model;

[0027] Second judgment:

[0028] Perform determination analysis on the first data set U1, the third data set U3 and the fourth data set U4 to obtain the corresponding tenth data set U10, the eleventh data set U11, the twelfth data set U12 and the thirteenth data set U13;

[0029] and use the tenth data set U10, the eleventh data set U11, the twelfth data set U12, the thirteenth data set U13 and the first identifier g1 as input information to train the residual network model;

[0030] Perform determination analysis on the second data set U2, the third data set U3, and the fifth data set U5 to obtain the corresponding fourteenth data set U14, the fifteenth data set U15, the sixteenth data set U16, and the seventeenth data set U17;

[0031] The fourteenth data set U14, the fifteenth data set U15, the sixteenth data set U16, the seventeenth data set U17 and the second identifier g2 are used as input information to train the residual network model.

[0032] Furthermore, in the residual network model, it includes:

[0033] Input layer: receives preprocessed image data as input of the model;

[0034] Several convolutional layers: use convolution kernels to extract features from input image data and perform convolution operations; standardize the input of each layer to keep it in the range of [0, 1];

[0035] Residual block: fuses the features extracted based on the shallow network model and the features extracted based on the deep network model; the features extracted based on the deep network model are the first feature results, and the features extracted based on the shallow network model are the second feature results;

[0036] Pooling layer: processes the feature map;

[0037] Fully connected layer: After extracting features, the obtained feature map is stretched into one-dimensional data and connected to the pooling layer. The feature vector output by the pooling layer is mapped to the category space. Each neuron is connected to all neurons in the previous layer and activated using a nonlinear activation function.

[0038] Output layer: Outputs the predicted value based on the classification results.

[0039] Furthermore, the process of responding to the feedback action is:

[0040] S101, determine whether there is an operation record for the distribution cabinet equipment currently being inspected:

[0041] When the distribution cabinet equipment under inspection has operation records, the historical operation records of the distribution cabinet equipment are obtained, and the abnormal state of the equipment components and the corresponding predicted values ​​are obtained within the predetermined time period t; a nonlinear correction calculation model is built to generate an abnormality degree evaluation value; otherwise, the abnormal state of the equipment components and the corresponding predicted values ​​are obtained within the predetermined time period t; a nonlinear correction calculation model is built to generate an abnormality degree evaluation value;

[0042] S102, comparing the abnormality evaluation value with a preset standard threshold, and generating different instructions according to the comparison result to execute corresponding strategies.

[0043] Further, in S101, the historical operation record of the power distribution equipment includes the number of component failures z1, the number of component failures z2, and the number of component maintenance times z3; the process of obtaining the abnormality evaluation value is as follows:

[0044] First, when it is judged to be an abnormal state, two time series are randomly extracted based on the corresponding moment of the abnormal state of the component, and the abnormal time period is calculated and generated. The abnormal time periods are arranged in time series and marked as T1, T2, T3, ..., Tn; where the time series is a dynamic change, and n represents the number of times it is judged to be an abnormal state;

[0045] Secondly, divide the abnormal time period into equal parts and mark them as m, where m is a dynamic value and when it has a remainder, it is m+1, where m is a positive integer; then retrieve the predicted value corresponding to the abnormal state, and calculate the abnormal factor yc in the corresponding abnormal time period according to m or m+1;

[0046] Finally, the abnormality evaluation value PG is generated according to the formula.

[0047] (III) Beneficial effects

[0048] The present invention provides a distribution cabinet equipment inspection system based on a residual network, which has the following beneficial effects:

[0049] 1. The present invention designs two methods: one is to extract features from the preprocessed image data based on a deep network model to obtain a first feature result; at the same time, the feature extraction from the preprocessed image data based on a shallow network model to obtain a second feature result; the residual network model is trained according to the first feature result and the second feature result; the extracted features have not only simple shallow network features but also more complex deep network features, thereby improving the integrity and accuracy of feature extraction;

[0050] 2. The residual network model proposed in the present invention involves a convolutional layer, a fully connected layer, a pooling layer, and a residual block, which can effectively solve the complexity of network depth and improve the efficiency of training; at the same time, based on the trained residual network model, the target state of the components of the distribution cabinet equipment can be analyzed more quickly to quickly distinguish the abnormal state for subsequent monitoring and early warning work;

[0051] 3. The present invention calculates the abnormality evaluation value of the components of the distribution cabinet equipment based on the historical operation records of the distribution cabinet equipment and the predicted values ​​trained by the residual network model, and builds a nonlinear correction calculation model. In this process, the abnormal time period is generated based on the analysis of the abnormal state moment to correct the model calculation, thereby achieving more accurate and intelligent status judgment of the components of the distribution cabinet equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A module schematic diagram of the present invention is shown. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Research and development concept: When inspecting the distribution cabinet equipment, the camera is first used to collect image data for model reasoning, specifically involving the RISP camera, and finally the classification results and image data are stored in the database;

[0055] Among them, the software involves data acquisition and preprocessing unit, deep analysis unit, state recognition unit and inspection feedback unit. By extracting features from image data and then training models in two ways, the classification of distribution cabinet equipment components is identified based on the trained residual network model to distinguish normal and abnormal states. Through real-time monitoring and data feedback, problems can be discovered and maintained in a timely manner to improve the safety and efficiency of equipment. Example

[0056] See also Figure 1, this embodiment provides a distribution cabinet equipment inspection system based on a residual network, the system includes various functional modules that run in sequence, namely, a data acquisition and preprocessing unit, a deep analysis unit, a state recognition unit and an inspection feedback unit, and the data acquisition and preprocessing unit, the deep analysis unit, the state recognition unit and the inspection feedback unit are communicatively connected; the application scenario of the system is: collecting image data of several distribution cabinets through a camera to identify the status of their equipment components, and perform monitoring and early warning processing;

[0057] Data acquisition and preprocessing unit:

[0058] Adopt image acquisition technology to collect image data of distribution cabinet equipment in real time and pre-process the image data;

[0059] The process of preprocessing image data is as follows:

[0060] Image recognition: Preset and build a device component library, analyze the collected image data, crop the target area of ​​the device component, and generate the device component image; the device component image is in RGB color format;

[0061] Color Model:

[0062] Grayscale processing: The device component image is converted into a grayscale image using the standard average method; according to the formula: ; Wherein, g is the grayscale image after grayscale processing;

[0063] Morphological operation: Based on the grayscale image, the opening operation is used to remove shadows; the opening operation is first corrosion operation and then expansion operation; according to the formula: ; Where A is the original device component image, B is the structural element, represents the corrosion operation, and ⊕ represents the expansion operation;

[0064] Binarization segmentation: After the opening operation, the binary classification is performed so that the gray value is 0 or 1;

[0065] In-depth analysis unit:

[0066] A residual network model is preset, in which a deep network model and a shallow network model are built-in; and feature extraction is performed on the preprocessed image data based on the deep network model to obtain a first feature result; and feature extraction is performed on the preprocessed image data based on the shallow network model to obtain a second feature result; and the residual network model is trained according to the first feature result and the second feature result to obtain an updated residual network model for the next feature extraction;

[0067] The process of obtaining the first feature result is:

[0068] Based on the deep network model, feature extraction is performed on the preprocessed image data to form feature vectors, and the feature vectors are concatenated to form a state feature matrix; the feature vectors include but are not limited to low-level features such as color features and texture features; and also include high-level features of object parts and object classification;

[0069] Matrix analysis: pre-set the feature matrix threshold, compare and analyze the state feature matrix corresponding to each device component image with the preset feature matrix threshold, and when the state feature matrix meets the preset feature matrix threshold, the corresponding position is recorded as 1, otherwise it is recorded as 0; and mark them in turn to form the first matrix r1;

[0070] Output result: Integrate the first matrix r1 and mark it as the first feature result;

[0071] At the same time, based on the shallow network model, feature extraction is performed on the image data to form a feature vector, and a second matrix r2 is constructed, and the second matrix r2 is integrated and marked as a second feature result; the feature vector includes low-level features such as color features and texture features;

[0072] For example: convert each extracted feature (such as color feature, edge feature, texture feature, etc.) into a feature vector. If HOG feature is used, a feature vector h=[h1, h2, h3…] is obtained, and different feature vectors are combined into a large feature vector f; organize the feature vectors into a matrix,

[0073] Usually each row represents an image sample and each column represents a feature to construct the second matrix ; And the second matrix r2 is integrated and marked as the second feature result;

[0074] The first feature result includes a first data set U1 obtained according to the deep network model that meets the preset standard, a second data set U2 that does not meet the preset standard, and a third data set U3 that cannot be judged whether it meets the preset standard; the second feature result includes a fourth data set U4 obtained according to the shallow network model that meets the preset standard and a fifth data set U5 that does not meet the preset standard; then the process of training the residual network model based on the first feature result and the second feature result is:

[0075] Marking identification: assigning corresponding identifications to the first data set U1, the second data set U2, the third data set U3, the fourth data set U4 and the fifth data set U5; wherein, the data that meet the preset standard is marked as the first identification g1, the data that do not meet the preset standard is marked as the second identification g2, and the data that cannot be judged whether it meets the preset standard is marked as the third identification g3; wherein, the first identification g1, the second identification g2 and the third identification g3 can be defined as hash table values ​​to facilitate data indexing;

[0076] The preset standard is a preset extraction feature template, and the preset extraction feature template also has built-in parameter settings for feature extraction; to determine whether it meets the preset standard, does not meet the preset standard, or cannot be determined whether it meets the preset standard;

[0077] Meet the preset criteria: the features can improve the performance of the model on the task;

[0078] For example, for edge features, use the extracted features to train the classifier and test the accuracy on the edge detection task. When the accuracy is high, it means that the features effectively capture the edge features and meet the preset standards.

[0079] Does not meet the preset standards: The features cause the model performance to deteriorate and cannot effectively separate different categories of data;

[0080] For example, for edge features, use the extracted features to train the classifier and test the accuracy on the edge detection task. If the accuracy is significantly lower than other methods, it means that the feature extraction does not meet the preset standards.

[0081] Unable to determine whether the preset standards are met: the requirements for the characteristics are unclear;

[0082] For example: the evaluation method of edge feature extraction is uncertain, which means that feature extraction cannot be judged;

[0083] Determination analysis:

[0084] First judgment: Perform judgment analysis on the first data set U1 and the fourth data set U4:

[0085] Acquire data that does not exist in the fourth data set U4 from the first data set U1 to obtain a sixth data set U6; acquire data that does not exist in the first data set U1 from the fourth data set U4 to obtain a seventh data set U7;

[0086] And use the sixth data set U6, the seventh data set U7 and the first identifier g1 as input information to train the residual network model;

[0087] The second data set U2 and the fifth data set U5 are also subjected to decision analysis:

[0088] Acquire data that does not exist in the fifth data set U5 in the second data set U2 to obtain an eighth data set U8; acquire data that does not exist in the second data set U2 in the fifth data set U5 to obtain a ninth data set U9;

[0089] And use the eighth data set U8, the ninth data set U9 and the second identifier g2 as input information to train the residual network model;

[0090] Second judgment:

[0091] Perform judgment analysis on the first data set U1, the third data set U3 and the fourth data set U4:

[0092] The data not included in the third data set U3 and the fourth data set U4 are obtained from the first data set U1 to obtain the tenth data set U10; the data not included in the first data set U1 and the third data set U3 are obtained from the fourth data set U4 to obtain the eleventh data set U11;

[0093] The data existing in the first data set U1 is obtained in the third data set U3 to obtain the twelfth data set U12; the data existing in the fourth data set U4 is obtained in the third data set U3 to obtain the thirteenth data set U13;

[0094] and use the tenth data set U10, the eleventh data set U11, the twelfth data set U12, the thirteenth data set U13 and the first identifier g1 as input information to train the residual network model;

[0095] The second data set U2, the third data set U3 and the fifth data set U5 are analyzed:

[0096] Acquire data from the second data set U2 that does not exist in the third data set U3 and the fifth data set U5 to obtain a fourteenth data set U14;

[0097] Acquire data that does not exist in the second data set U2 and the third data set U3 in the fifth data set U5 to obtain a fifteenth data set U15;

[0098] The data existing in the second data set U2 is obtained in the third data set U3 to obtain the sixteenth data set U16; the data existing in the fifth data set U5 is obtained in the third data set U3 to obtain the seventeenth data set U17;

[0099] And use the fourteenth data set U14, the fifteenth data set U15, the sixteenth data set U16, the seventeenth data set U17 and the second identifier g2 as input information to train the residual network model;

[0100] It should be noted that, in this embodiment, the integrity of feature extraction is ensured by combining the above two methods for training. At the same time, as the training process continues, the preset feature extraction template is updated to achieve optimal adaptability.

[0101] For example: select edge features and texture features for analysis:

[0102] The template corresponding to the edge feature is:

[0103] Sobel operator: kernel size;

[0104] Canny operator: low threshold, high threshold, kernel size;

[0105] Template corresponding to texture features:

[0106] Gray-level co-occurrence matrix: distance and direction angle between pixel pairs;

[0107] Local binary pattern: the number of neighborhood pixels, radius, method type, etc.; change the parameters of edge features and texture features during training;

[0108] State recognition unit:

[0109] Based on the residual network model, several convolutional layers are used to extract features, and the features extracted by the shallow network model and the deep network model are fused through the residual block. The feature dimension is then processed through the pooling layer. Finally, the fully connected layer is used to map the extracted features to the category space, and the output layer is used to output the predicted value.

[0110] Among them, one round of training cycle is composed of one judgment and two judgments, and the learning rate of the residual network model is set to 0.1 and the batch size is 512 to prevent overfitting; and starting from the second round of training cycle, the shallow network model is eliminated after at least 3*40 rounds; in this process, the residual block is introduced to ensure the accuracy and precision of the residual network model in the process of deepening the network depth;

[0111] In the residual network model, including:

[0112] Input layer: receives preprocessed image data as input of the model;

[0113] For example: Assume that the size of the input image data is 3*224*224, that is, 3 channels, and the size of each channel is 224*224;

[0114] Several convolutional layers: use convolution kernels to extract features from input image data and perform convolution operations;

[0115] For example, the settings are:

[0116] The first 7*7 convolution layer: the size of the convolution kernel is 7*7 convolution layer, the step size is 2, the padding is 3, the output channel is 64, then the data size is 64*112*112;

[0117] The first 3*3 convolution layer: the size of the convolution kernel is 3*3, the step size is 1, and the padding is 1; and it is set to 2, then the final output data size is 64*56*56, that is, this layer does not change the data size and the number of channels;

[0118] The second 3*3 convolution layer: passes through a 1*1 convolution layer and undergoes a downsampling; the final output data is 128*28*28, which doubles the output channel and halves the output data size;

[0119] The third 3*3 convolution layer: the same 1*1 convolution layer and downsampling are performed, and the final output is 256*14*14, which doubles the output channels and halves the output data size;

[0120] The fourth 3*3 convolution layer: the same 1*1 convolution layer and downsampling are performed, and the final output is 512*7*7, which doubles the output channels and halves the output data size;

[0121] At the same time, the input of each layer is standardized to be controlled in the interval [0, 1]. It should be noted that, assuming that the weights used in the convolution operation form a convolution kernel of size b*b, and b is a positive integer, the grayscale value of the pixel on the image is multiplied by the value on the corresponding convolution kernel, and then the multiplication result is used as the grayscale value of the pixel on the image corresponding to the middle pixel of the convolution kernel;

[0122] Residual block: fuses the features extracted based on the shallow network model and the features extracted based on the deep network model; the features extracted based on the deep network model are the first feature results, and the features extracted based on the shallow network model are the second feature results;

[0123] It should be noted that the core idea of ​​the residual block is to introduce shortcut connections to add the input directly to the output, thereby alleviating the gradient vanishing problem of the deep network and improving the training effect; for example: shallow network features extracted based on shallow network models, including edge and texture features; features extracted based on deep network models, including edge, texture, object parts or category features, etc.;

[0124] Pooling layer: Process the feature map, including reducing the size and dimension, which is achieved through pooling operation;

[0125] It should be noted that common pooling operations include maximum pooling and average pooling; maximum pooling: use a 2*2 pooling window with a step size of 2, and select the maximum value in each region as the output; average pooling: use a 2*2 pooling window with a step size of 2, and calculate the average value in each region as the output;

[0126] For example: if the maximum pooling is used on the output of the second convolutional layer, the size of the output data is 64*56*56; if the average pooling is used on the output of the last convolutional layer, the final output data size is 512*1*1;

[0127] Fully connected layer: After extracting features, the obtained feature map is stretched into one-dimensional data and connected to the pooling layer. The feature vector output by the pooling layer is mapped to the category space. Each neuron is connected to all neurons in the previous layer and activated using a nonlinear activation function to improve the nonlinear fitting ability of the model.

[0128] Output layer: outputs the predicted value according to the classification results;

[0129] For example: use the softmax function to calculate the probability distribution of each sample, for each sample, calculate the cross entropy loss of the true category, and select the smallest one as the final prediction value;

[0130] Inspection feedback unit:

[0131] The target state of the components of the power distribution cabinet is determined based on the predicted value. The target state includes a normal state or an abnormal state. When it is determined to be a normal state, no response feedback action is taken; when it is determined to be an abnormal state, a response feedback action is taken;

[0132] Among them, a display terminal is also provided to display the target status of the components of the power distribution cabinet;

[0133] First, there are 9 types of components, including energy storage state, pressure plate, switch, button, 2 types of knobs, arrow, indicator, and indicator light. Each component has 2 or 3 states, which are described in the format of "component type_state":

[0134] Energy storage status_energy stored, energy storage status_energy not stored;

[0135] Press plate_oblique, press plate_vertical;

[0136] switch_left, switch_right;

[0137] Button_on, Button_off;

[0138] Knob 1_Left, Knob 1_Middle, Knob 1_Right;

[0139] Knob 2_up, Knob 2_middle, Knob 2_down;

[0140] arrow_up, arrow_down;

[0141] Indicator_oblique, indicator_horizontal, indicator_off, indicator_vertical;

[0142] Indicator light_on, indicator light_off; 22 states in total;

[0143] Then, when an abnormal state occurs, a response feedback action is immediately performed, and the process is as follows:

[0144] S101, determine whether there is an operation record for the distribution cabinet equipment currently being inspected:

[0145] When the distribution cabinet equipment under inspection has operation records, the historical operation records of the distribution cabinet equipment are obtained, and the abnormal state of the equipment components and the corresponding predicted values ​​are obtained within the predetermined time period t; a nonlinear correction calculation model is built to generate an abnormality degree evaluation value; otherwise, the abnormal state of the equipment components and the corresponding predicted values ​​are obtained within the predetermined time period t; a nonlinear correction calculation model is built to generate an abnormality degree evaluation value;

[0146] The historical operation records of the power distribution equipment include the number of component failures z1, the number of component failures z2, and the number of component maintenance times z3; the process of obtaining the abnormality evaluation value is as follows:

[0147] First, when it is judged to be an abnormal state, two time series are randomly extracted based on the corresponding moment of the abnormal state of the component, and the abnormal time period is calculated and generated. The abnormal time periods are arranged in time series and marked as T1, T2, T3, ..., Tn; wherein the time series is a dynamic change quantity, and n represents the number of times it is judged to be an abnormal state; for example: a time series t1 is randomly extracted forward based on the moment of obtaining the abnormal state, and a time series t2 is randomly extracted backward based on the moment of the abnormal state, and a time period is formed based on the time series t1 and the time series t2; then other time periods are obtained in sequence, and the time periods are sorted according to the time series;

[0148] Secondly, divide the abnormal time period into equal parts and mark them as m, where m is a dynamic value and when it has a remainder, it is m+1, where m is a positive integer; at the same time, retrieve the prediction value corresponding to the abnormal state; for example, if the time between the start time of the abnormal time period and the abnormal time is 5s, if it is divided into 2s, 5 divided by 2, the remainder is 1, and the value of m is 6;

[0149] When it is the abnormal time period corresponding to m, according to the formula: Calculate the abnormal factor yc in the corresponding abnormal time period;

[0150] When it is the abnormal time period corresponding to m+1, according to the formula: Calculate the abnormal factor yc in the corresponding abnormal time period; 1. Where T represents the time period, and the value is T1, T2..., Tn, and the calculation method is the length of time between the start time and the end time, It is the average value of the predicted values ​​collected in any abnormal time period. α1 and α2 are weight correction coefficients, and both α1 and α2 are greater than 0;

[0151] Finally, the formula used to generate the anomaly assessment value is as follows:

[0152] ;

[0153] Among them, PG is the abnormality evaluation value, p1 indicates that there is no historical operation record, p2 indicates that there is a historical operation record, β1, β2 and β3 are all weight correction coefficients, and β1, β2 and β3 are all greater than 0; it should be noted that the larger the abnormality evaluation coefficient is, the more obvious the abnormal state of the equipment component is, and the greater the safety hazard is;

[0154] S102, comparing the abnormality evaluation value PG with a preset standard threshold, and generating different instructions according to the comparison result to execute corresponding strategies; wherein the preset standard threshold is a preset threshold stored in the system, and its value is set based on the statistical characteristics of the data, by obtaining the mean and standard deviation corresponding to the abnormal evaluation value within a preset time period t, and setting the standard threshold to the mean plus twice the standard deviation;

[0155] The process of executing the corresponding strategy is:

[0156] When the abnormality evaluation value PG is less than the standard threshold, it means that the abnormality of the corresponding equipment component is low, and a monitoring instruction is generated and displayed, with the text content "the abnormality is low, and the monitoring modification strategy needs to be executed". Professionals mark the equipment component and include it in the periodic monitoring plan, and perform inspections and maintenance according to the established cycle (such as one week) to prevent the potential abnormal problems from worsening;

[0157] When the abnormality assessment value PG ≥ the standard threshold, it means that the abnormality of the corresponding equipment component is high, and an early warning instruction is generated and displayed. The text content is "The abnormality is high and an immediate early warning strategy is required". Professional personnel will perform maintenance and replacement operations to prevent other problems from occurring in the distribution cabinet.

[0158] In summary, the present invention includes a data acquisition and preprocessing unit, a deep analysis unit, a state recognition unit and an inspection feedback unit. It acquires images of distribution cabinet equipment, and performs feature extraction on the preprocessed image data based on a deep network model to obtain a first feature result, and performs feature extraction on the preprocessed image data based on a shallow network model to obtain a second feature result. At the same time, a residual network model is trained based on the first feature result and the second feature result to ensure that the collected features are complete and accurate enough to reduce detection errors. The residual network model involves convolutional layers, fully connected layers, pooling layers and residual blocks, which can effectively solve the complexity of network depth, improve training efficiency, and analyze the target state of distribution cabinet equipment components more quickly to distinguish abnormal states for subsequent monitoring and early warning.

[0159] The size of the interval and threshold is set for the convenience of comparison. The size of the threshold depends on the amount of sample data and the number of bases set by technicians in this field for each group of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value;

[0160] The above formulas are all dimensionless and numerical calculations. The formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0161] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0162] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. Power distribution cabinet equipment inspection system based on residual network, Features: include: The data acquisition and preprocessing unit uses image acquisition technology to collect image data of the power distribution cabinet equipment in real time and preprocess the image data; A deep analysis unit is configured to preset a residual network model, wherein a deep network model and a shallow network model are built in the residual network model; and based on the deep network model, feature extraction is performed on the preprocessed image data to obtain a first feature result; and at the same time, based on the shallow network model, feature extraction is performed on the preprocessed image data to obtain a second feature result; And training the residual network model according to the first feature result and the second feature result to obtain an updated residual network model for next feature extraction; Among them, the first feature result includes a first data set U1 obtained according to the deep network model that meets the preset standard, a second data set U2 that does not meet the preset standard, and a third data set U3 that cannot be judged whether it meets the preset standard; the second feature result includes a fourth data set U4 obtained according to the shallow network model that meets the preset standard and a fifth data set U5 that does not meet the preset standard; then the process of training the residual network model based on the first feature result and the second feature result is: Marking identification: assigning corresponding identifications to the first data set U1, the second data set U2, the third data set U3, the fourth data set U4 and the fifth data set U5; wherein, the data that meet the preset standard are marked as the first identification g1, the data that do not meet the preset standard are marked as the second identification g2, and the data that cannot be judged whether it meets the preset standard are marked as the third identification g3; Determination analysis: First judgment: perform judgment analysis on the first data set U1 and the fourth data set U4 to obtain the corresponding sixth data set U6 and seventh data set U7; and use the sixth data set U6, the seventh data set U7 and the first identifier g1 as input information to train the residual network model; The second data set U2 and the fifth data set U5 are subjected to judgment analysis to obtain the corresponding eighth data set U8 and ninth data set U9; and the eighth data set U8, the ninth data set U9 and the second identifier g2 are used as input information to train the residual network model; Second judgment: Perform determination analysis on the first data set U1, the third data set U3 and the fourth data set U4 to obtain the corresponding tenth data set U10, the eleventh data set U11, the twelfth data set U12 and the thirteenth data set U13; and use the tenth data set U10, the eleventh data set U11, the twelfth data set U12, the thirteenth data set U13 and the first identifier g1 as input information to train the residual network model; Perform determination analysis on the second data set U2, the third data set U3, and the fifth data set U5 to obtain the corresponding fourteenth data set U14, the fifteenth data set U15, the sixteenth data set U16, and the seventeenth data set U17; And use the fourteenth data set U14, the fifteenth data set U15, the sixteenth data set U16, the seventeenth data set U17 and the second identifier g2 as input information to train the residual network model; The state recognition unit is based on the residual network model. It extracts features through several convolutional layers, and fuses the features extracted by the shallow network model with the features extracted by the deep network model through the residual block. Then, the feature dimension is processed through the pooling layer. Finally, the fully connected layer is used to map the extracted features to the category space, and the output layer is used to output the predicted value. The inspection feedback unit determines the target state of the distribution cabinet equipment components based on the predicted value. The target state includes a normal state or an abnormal state. When it is determined to be a normal state, no response feedback action is taken; when it is determined to be an abnormal state, a response feedback action is taken.

2. According to the residual network-based distribution cabinet equipment inspection system of claim 1, Features: The process of preprocessing image data is as follows: Image recognition: Preset and build a device component library, analyze the collected image data, crop the target area of ​​the device component, and generate the device component image; the device component image is in RGB color format; Color model: Grayscale processing: Use the average method to convert the device component image into a grayscale image; Morphological operation: Use the opening operation based on the grayscale image to remove shadows; Binarization segmentation: After the opening operation, perform binary classification so that its grayscale value is 0 or 1.

3. According to the residual network-based distribution cabinet equipment inspection system of claim 1, Features: The process of obtaining the first feature result is as follows: Feature extraction: Extract features from the preprocessed image data based on a deep network model to form feature vectors, and concatenate the feature vectors to form a state feature matrix; feature vectors include but are not limited to color features and texture features; Matrix analysis: pre-set the feature matrix threshold, compare and analyze the state feature matrix corresponding to each device component image with the preset feature matrix threshold, and when the state feature matrix meets the preset feature matrix threshold, the corresponding position is recorded as 1, otherwise it is recorded as 0; and mark them in turn to form the first matrix r1; Output result: Integrate the first matrix r1 and mark it as the first feature result.

4. According to the residual network-based distribution cabinet equipment inspection system of claim 1, Features: Based on the shallow network model, feature extraction is performed on the image data to form a feature vector, and a second matrix r2 is constructed, and the second matrix r2 is integrated and marked as a second feature result.

5. According to the residual network-based distribution cabinet equipment inspection system of claim 1, Features: In the residual network model, including: Input layer: receives preprocessed image data as input of the model; Several convolutional layers: use convolution kernels to extract features from input image data and perform convolution operations; standardize the input of each layer to keep it in the range of [0, 1]; Residual block: fuses the features extracted based on the shallow network model and the features extracted based on the deep network model; the features extracted based on the deep network model are the first feature results, and the features extracted based on the shallow network model are the second feature results; Pooling layer: processes the feature map; Fully connected layer: After extracting features, the obtained feature map is stretched into one-dimensional data and connected to the pooling layer. The feature vector output by the pooling layer is mapped to the category space. Each neuron is connected to all neurons in the previous layer and activated using a nonlinear activation function. Output layer: Outputs the predicted value based on the classification results.

6. According to the residual network-based distribution cabinet equipment inspection system of claim 1, Features: The process of responding to feedback actions is: S101, determine whether there is an operation record for the distribution cabinet equipment currently being inspected: When the distribution cabinet equipment under inspection has an operation record, the historical operation record of the distribution cabinet equipment is obtained, and the abnormal state of the equipment components and the corresponding prediction value are obtained in a predetermined time period t; Build a nonlinear correction calculation model to generate anomaly assessment values; Otherwise, the abnormal state of the equipment component and the corresponding predicted value are obtained in the predetermined time period t; Build a nonlinear correction calculation model to generate anomaly assessment values; S102, comparing the abnormality evaluation value with a preset standard threshold, and generating different instructions according to the comparison result to execute corresponding strategies.

7. According to claim 6, the residual network-based distribution cabinet equipment inspection system, Features: In S101, the historical operation records of the power distribution equipment include the number of component failures z1, the number of component failures z2, and the number of component maintenance times z3; the process of obtaining the abnormality evaluation value is as follows: First, when it is judged to be an abnormal state, two time series are randomly extracted based on the corresponding moment of the abnormal state of the component, and the abnormal time period is calculated and generated. The abnormal time periods are arranged in time series and marked as T1, T2, T3, ..., Tn; where the time series is a dynamic change, and n represents the number of times it is judged to be an abnormal state; Secondly, divide the abnormal time period into equal parts and mark them as m, where m is a dynamic value and when it has a remainder, it is m+1, where m is a positive integer; then retrieve the predicted value corresponding to the abnormal state, and calculate the abnormal factor yc in the corresponding abnormal time period under m or m+1 respectively; When it is the abnormal time period corresponding to m, according to the formula: Calculate the abnormal factor yc in the corresponding abnormal time period; When it is the abnormal time period corresponding to m+1, according to the formula: Calculate the abnormal factor yc in the corresponding abnormal time period; 1. Where T represents the time period, and the value is T1, T2..., Tn, and the calculation method is the length of time between the start time and the end time, It is the average value of the predicted values ​​collected in any abnormal time period. α1 and α2 are weight correction coefficients, and both α1 and α2 are greater than 0; Finally, according to the formula: Generate an abnormality evaluation value PG; wherein PG is an abnormality evaluation value, p1 indicates that there is no historical operation record, p2 indicates that there is a historical operation record, β1, β2 and β3 are all weight correction coefficients, and β1, β2 and β3 are all greater than 0.

Citation Information

Patent Citations

  • Insulator state detection method based on adaptive optimization of residual offset network

    CN117474883A

  • Distribution line unmanned aerial vehicle inspection image target detection and hidden danger intelligent identification method

    CN117726958A

  • Auxiliary welding device for electric control cabinet frame

    CN220782751U

  • Myocardial fibrosis classification method based on residual capsule network

    WO2024108505A1

  • KR20230016470A