A device fault diagnosis method based on deep residual network

Through the method based on deep residual network, the time domain characteristics of the power transmission and transformation equipment are extracted and the model is trained, and the problem of insufficient evaluation in the existing technology is solved, precise identification and fault diagnosis of the power transmission and transformation equipment status is realized, and the evaluation is improved. accuracy and practicality.

CN114091549BActive Publication Date: 2024-12-03STATE GRID JIANGSU ELECTRIC POWER CO LTD SUZHOU BRANCH +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202111145014.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-28
Publication Date
2024-12-03
Estimated Expiration
2041-09-28

AI Technical Summary

Technical Problem

When evaluating the status of power transmission and transformation equipment, the existing technology lacks multi-link and multi-dimensional considerations, resulting in insufficient accuracy and targeting of the evaluation results. Due to insufficient data on the fault sample, the key parameters of the evaluation model mainly rely on empirical selection, which is seriously restricted. The accuracy and practicality of the evaluation results were carried out.

Method used

By using a deep residual network method, the operation status data of the transmission and transformation equipment is collected, the time domain features are extracted, the training set and test set are constructed, and the deep residual network model is trained to achieve differentiation of the status of the transmission and transformation equipment and Fine identification.

Benefits of technology

It realizes accurate identification and fault diagnosis of the status of power transmission and transformation equipment, improves the accuracy and practicality of evaluation, and provides dispatchers with more accurate equipment information support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114091549B_ABST
    Figure CN114091549B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for fault-tolerant diagnosis of equipment faults based on a deep residual network, the method comprising the following steps: collecting the operating status data of power transmission and transformation equipment; extracting feature values ​​of the collected status data at certain time intervals to obtain the time domain features of each device under different operating states; classifying and encoding the extracted time domain feature data according to the equipment type and operating state; constructing a training set and a test set, inputting the classified and encoded data into a deep residual network to train a deep network model; and using the trained deep residual network model to perform differentiated and refined identification of the status of power transmission and transformation equipment. The present invention collects the operating parameters of power transmission and transformation equipment, extracts the time domain waveforms of key parameters, and uses a deep residual network to train the time domain characteristics of each device under different operating states, and compares them with the input time domain features to achieve equipment status identification and fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of power system equipment fault diagnosis, and specifically relates to an equipment fault diagnosis method based on a deep residual network. Background Art

[0002] Accurately assessing the operating status of key power transmission and transformation equipment and formulating targeted operation and maintenance strategies in a timely manner are of great significance to ensuring the safe, reliable and economical operation of the power system. With the rapid development of smart grid technology, information technology and intelligent operation and maintenance technology, the status data of power transmission and transformation equipment has gradually shown big data characteristics such as large volume, multiple types and rapid growth. How to achieve efficient use of multivariate data, deeply explore the relationship between various types of data and equipment status, and realize refined evaluation of equipment status has become the key research content of power transmission and transformation equipment fault diagnosis in power systems.

[0003] Most of the existing state assessment methods based on equipment test data are based on single or limited state parameters, lacking multi-link and multi-dimensional considerations, and the accuracy and pertinence of the assessment results need to be improved. In addition, since defects or failures are low-probability events, the existing defect and failure sample data cannot meet the needs of intelligent assessment methods for modeling samples, resulting in the key parameters of the assessment model mainly relying on experience selection, which seriously restricts the accuracy of the assessment results and the practicality of the assessment method. Therefore, it is urgent to explore new state assessment methods, establish a state assessment model that comprehensively considers the influence of external factors, and improve the accuracy of the assessment. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the purpose of the present invention is to provide an equipment fault diagnosis method based on a deep residual network, and to train and verify the deep residual network model based on the historical data of the operating status of the power transmission and transformation equipment, so as to realize the differentiated and refined evaluation of the operating status of the power transmission and transformation equipment, and to achieve accurate identification of the status of the power transmission and transformation equipment.

[0005] The present invention adopts the following technical solution:

[0006] A device fault tolerance diagnosis method based on a deep residual network, the method comprising the following steps:

[0007] Step 1: Collect the operating status data of power transmission and transformation equipment;

[0008] Step 2: Extract feature values ​​from the collected status data at certain time intervals to obtain the time domain features of each device under different operating conditions;

[0009] Step 3: Classify and encode the extracted time domain feature data according to the equipment type and operating status;

[0010] Step 4: Construct training sets and test sets, and input the classified and encoded data into the deep residual network to train the deep network model;

[0011] Step 5: Use the trained deep residual network model to differentiate and refine the status of power transmission and transformation equipment.

[0012] Preferably, in step 1, the status data is derived from equipment data, monitoring data and external data;

[0013] Equipment data includes nameplate parameters, ledger data, operation data, inspection data, defect / elimination data, and fault / maintenance records;

[0014] The detection data includes live detection data and online monitoring data; live monitoring data includes infrared detection, ultraviolet detection, and ultrasonic detection; online monitoring data includes dissolved gas in oil, partial discharge, core grounding current, and gas pressure;

[0015] External data includes meteorology, environment, geography, power grid dispatch, and load.

[0016] Preferably, feature value extraction is performed on the collected state parameters at certain time intervals to obtain time domain features of each device under different operating states;

[0017] By comparing the importance of different types of state quantities in the data set, irrelevant and redundant attributes are deleted to extract key state parameters;

[0018] After selecting the key parameters, the time domain characteristics of each device under different operating conditions are obtained according to the time domain analysis method.

[0019] Preferably, in step 3, the extracted time domain feature data is classified according to the device type and the operating status, and different types of data are encoded and labeled accordingly.

[0020] Preferably, the equipment types include power transformers, high voltage circuit breakers, disconnectors, voltage / current transformers, and transmission lines;

[0021] The operating status includes normal status and different fault status; the fault status of the power transformer includes overheating fault, winding fault, and discharge fault; the fault status of the high-voltage circuit breaker includes mechanical fault, secondary circuit fault, and sealing failure fault, etc.; the fault status of the disconnector includes overheating fault, mechanical fault, and insulation fault; the fault status of the voltage / current transformer includes mechanical fault, insulation fault, secondary circuit fault, and resonance fault, and the fault status of the transmission line includes grounding fault, circuit breaker fault, open circuit fault, and overload fault.

[0022] Preferably, in step 4, 80% of the time domain feature data of different operating states of each device are randomly selected as training sets and validation sets to perform parameter training on the deep residual network model; the remaining 20% ​​are used as a test set to analyze and test the performance of the deep residual network model.

[0023] Preferably, in step 4, the deep residual network includes residual modules, and multiple residual modules are connected end to end to form a deep residual network.

[0024] Preferably, in step 4, the residual module consists of a convolutional layer, a batch normalization layer and an activation function layer.

[0025] Preferably, the convolution layer uses multiple different convolution kernels to extract the main features of the input data. For one-dimensional input data, the output of the lth convolution layer is:

[0026]

[0027] in, is the output of the lth convolutional layer, is the input of the lth convolutional layer, is the convolution kernel weight, is the corresponding bias.

[0028] Preferably, batch normalization is performed so that the outputs of each layer conform to a Gaussian distribution with the same mean and variance. For a small batch of data x i , the operation of batch normalization layer is expressed as:

[0029]

[0030] Among them, μ is the mean of a set of data, ε is the minimum value, and σ is the variance of the set of data. The data after one layer of batch normalization is processed.

[0031] Preferably, the activation function layer adopts the Relu function, which is expressed as:

[0032] Relu(x)=max(0,x).

[0033] Preferably, the final output of the residual module is:

[0034] y=x+F(x,w c )

[0035] Where x is the residual block input, F is a nonlinear function including batch normalization layer, activation function layer and convolution layer, and w c The parameters that need to be optimized for this path.

[0036] Preferably, the deep residual network includes 9 convolutional layers, wherein the first convolutional layer adopts a wide convolution kernel, and the remaining convolutional layers are composed of 4 residual modules, each of which includes 2 convolutional layers;

[0037] A global average pooling layer is used after the residual module, and a Dropout layer is added after the global average pooling layer. Finally, fault classification is achieved through the Softmax classifier.

[0038] Preferably, in step 4, the extracted time domain feature data is divided into a training set, a validation set and a test set in a ratio of 6:2:2;

[0039] Input the training set and validation set into the constructed deep residual network in batches, and calculate the error between the predicted value and the actual result;

[0040] After each round of iteration of the training set, the validation set is used for evaluation once;

[0041] The weights of each layer of the DRN network are updated according to the average cross entropy algorithm, and the above process is repeated until the loss function is minimized to obtain the optimal deep residual network;

[0042] Finally, the test set is input into the trained deep residual network, and the diagnosis results are output to verify the diagnostic ability of the model.

[0043] Preferably, the loss function adopts the Softmax loss function.

[0044] The function of Softmax is a classifier. Given an input prediction value x i And the corresponding true label value y i , and its calculation formula is:

[0045]

[0046] Among them, x i represents the input feature of the i-th one; y i Indicates the label of the category corresponding to the i-th input, Represents the bias corresponding to the true category of the i-th sample, W j represents the weight vector of the jth class, represents the weight vector of the input corresponding to the category label, T represents the vector transpose, b j Represents the bias of the sample category, and n represents the number of sample categories;

[0047] The calculation formula of Softmax loss is:

[0048]

[0049] Among them, L srepresents the Softmax loss function, that is, the sample average cross entropy; m represents the number of training samples.

[0050] The beneficial effect of the present invention is that, compared with the prior art, the deep residual network model is trained and verified based on the historical data of the operating status of the power transmission and transformation equipment, which can analyze the operating parameters of the power transmission and transformation equipment, accurately evaluate its operating status, and realize accurate identification of the status of the power transmission and transformation equipment, providing support for dispatchers to grasp the information of the power transmission and transformation equipment. The present invention collects the operating parameters of the power transmission and transformation equipment, extracts the time domain waveform of the key parameters, and uses the deep residual network to train the time domain characteristics of each device under different operating conditions, and compares them with the input time domain characteristics to realize the state identification and fault diagnosis of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the steps of the device fault tolerance diagnosis method based on the deep residual network of the present invention;

[0052] Figure 2 It is a more detailed step flow diagram of the device fault tolerance diagnosis method based on deep residual network of the present invention;

[0053] Figure 3 is a schematic diagram of a typical residual unit of a deep residual network of the present invention;

[0054] Figure 4 It is a schematic diagram of the deep residual network model structure of the present invention. DETAILED DESCRIPTION

[0055] The present application is further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present application.

[0056] like Figure 1 and Figure 2 As shown, the device fault tolerance diagnosis method based on deep residual network of the present invention comprises the following steps:

[0057] Step 1: Collect the operating status data of power transmission and transformation equipment.

[0058] Status data is the basis for evaluating the operating status of power transmission and transformation equipment, and it mainly comes from three aspects: equipment data, monitoring data, and external data. Equipment data includes nameplate parameters, ledger data, operating data, inspection data, defect / elimination data, and fault / maintenance records. Detection data includes live detection data and online monitoring data. Live monitoring data includes infrared detection, ultraviolet detection, ultrasonic detection, etc. Online monitoring data includes dissolved gas in oil, partial discharge, core grounding current, gas pressure, etc. External data includes meteorology, environment, geography, power grid dispatch, load, etc. The sources and specific types of status data are shown in Table 1.

[0059] Table 1 Sources of status data

[0060]

[0061] Step 2: Extract feature values ​​from the collected status data at certain time intervals to obtain the time domain features of each device under different operating conditions.

[0062] The collected state parameters are subjected to feature value extraction at certain time intervals to obtain the time domain characteristics of each device under different operating states. In the present invention, the time interval can be flexibly adjusted according to specific requirements. During the data collection process, the amount of data acquired can be adjusted by adjusting the collection time interval. By comparing the importance of different types of state quantities in the data set, irrelevant attributes and redundant attributes are deleted, and key state parameters are extracted. Commonly used feature selection methods include grey correlation analysis, correlation coefficient method, association rule mining, etc. For example, according to the association rule mining between the fault type and the state parameter of the transformer, its key state parameters under different states are shown in Table 2:

[0063] Table 2 Key state parameters of some transformer failure modes

[0064]

[0065] After selecting the key parameters, the time domain characteristics of each device under different operating states are obtained according to the time domain analysis method. Time domain analysis uses time as a measure to obtain a curve of each state parameter changing over time. This time domain waveform will be used as the input for fault diagnosis and compared with the waveforms of each operating state obtained after deep learning training to identify the corresponding fault type. Step 3: Classify and encode the extracted time domain feature data according to the device type and operating state.

[0066] The extracted time domain feature data is classified and coded according to the equipment type and operating status. The collected data is classified according to the equipment type and operating status, and different types of data are coded and labeled accordingly.

[0067] Among them, the equipment types include power transformers, high-voltage circuit breakers, disconnectors, voltage / current transformers, and transmission lines, etc., and the operating status includes normal status and different fault status. The fault status of different equipment may be different. For example, the fault status of power transformers includes overheating fault, winding fault, discharge fault, etc., the fault status of high-voltage circuit breakers includes mechanical fault, secondary circuit fault, sealing failure fault, etc., the fault status of disconnectors includes overheating fault, mechanical fault, insulation fault, etc., the fault status of voltage / current transformers includes mechanical fault, insulation fault, secondary circuit fault, resonance fault, etc., and the fault status of transmission lines includes grounding fault, circuit breaker fault, circuit breaker fault, overload fault, etc.

[0068] The equipment types and typical fault status classification of power transmission and transformation equipment are shown in Table 3:

[0069] Table 3 Equipment types and typical fault status classification of power transmission and transformation equipment

[0070]

[0071] Step 4: Construct training sets and test sets, and input the classified and encoded data into the deep residual network for model training.

[0072] Construct training sets and test sets, input the classified and encoded data into the Deep Residual Network (DRN) for model training and verification, find the relationship between time domain feature data and equipment operation and operating status, so that the operating status can be judged based on the time domain feature data.

[0073] 80% of the time domain feature data of different operating states of each device are randomly selected as training sets and validation sets to train the parameters of the DRN model; the remaining 20% ​​are used as test sets to analyze and test the performance of the DRN model. That is, the input data of the deep residual network is the time domain curve processed in step 3.

[0074] The deep residual network model is built on the basis of convolutional neural network (CNN), and solves the gradient vanishing problem of deep network structure by adding residual modules.

[0075] The residual module is the core of the deep residual network. Multiple residual modules are connected end to end to build a deep residual network. The residual module consists of a convolutional layer, a batch normalization layer (BN layer) and an activation function layer. Its structure is as follows: Figure 3 shown.

[0076] The convolution layer uses multiple different convolution kernels to extract the main features of the input data. For one-dimensional input data, the convolution process is as follows, and the output of the lth convolution layer is:

[0077]

[0078] In the formula, is the output of the lth convolutional layer, is the input of the lth convolutional layer, is the convolution kernel weight, is the corresponding bias.

[0079] The BN layer makes the output of each layer conform to the Gaussian distribution with the same mean and variance, making the data more stable and accelerating training. i , the operation of BN can be expressed as:

[0080]

[0081] In the formula, μ is the mean of a set of data, ε is the minimum value, and σ is the variance of the set of data. is the data after BN processing.

[0082] Relu is a commonly used activation function, which can be expressed as:

[0083] Relu(x)=max(0,x)

[0084] The final output of the residual module is:

[0085] y=x+F(x,w c )

[0086] Where x is the residual block input, F is a nonlinear function including BN layer, activation function layer and convolution path, and w c The parameters that need to be optimized for this path.

[0087] The deep residual network model structure constructed by the present invention is as follows Figure 4 As shown in the figure. This model has 9 convolutional layers in total, of which the first convolutional layer uses a wide convolution kernel, and the remaining convolutional layers are composed of 4 residual modules, each of which contains 2 convolutional layers. A global average pooling layer is used after the residual module to reduce the amount of calculation. At the same time, in order to suppress overfitting, a Dropout layer is added after the global average pooling layer, and finally fault classification is achieved through the Softmax classifier.

[0088] The pooling layer, also known as the downsampling layer, performs downsampling and dimensionality reduction operations in this layer. Its main function is to reduce the number of data and parameters, reduce the computational load, effectively control overfitting, and make the input data "translation invariant". Common pooling methods include: max pooling, mean pooling, stochastic pooling, mixed pooling, etc. Max pooling takes the maximum value in the sampling area, mean pooling takes the average value in the sampling area, stochastic pooling takes values ​​randomly with a certain probability in the sampling area, and mixed pooling is a combination of max pooling and mean pooling. Among them, the formula for average pooling is:

[0089]

[0090] The formula for maximum pooling is:

[0091]

[0092] in, is the pooling result, represents the (i, j)th element in the kth feature map output by the lth convolutional layer, and p and q represent the length and width of the pooling window, respectively.

[0093] Global average pooling is to set the length and width of the average pooling window to the length and width of the input data window, and all data are calculated. The present invention uses global average pooling instead of the fully connected layer.

[0094] In addition, during deep learning training, if the training samples are insufficient, overfitting is prone to occur. Overfitting will reduce the actual prediction effect. The concept of the Dropout layer was proposed in 2012. Its essence is that during training, neurons become 0 with probability P and do not participate in the forward propagation process, which reduces the model's dependence on local features, so the model has better generalization ability. After experimental verification, the dropout layer can not only effectively reduce the overfitting phenomenon, but also reduce the prediction error rate a lot.

[0095] The extracted time domain feature data is divided into training set, validation set and test set in a ratio of 6:2:2. The training set and validation set are input into the constructed DRN network in batches to calculate the error between the predicted value and the actual result. After each round of iteration of the training set, the validation set is used for evaluation. The weights of each layer of the DRN network are updated according to the average cross entropy algorithm, and the above process is repeated until the loss function is minimized to obtain the optimal deep residual network model. Finally, the test set is input into the trained deep residual network, and the diagnosis results are output to verify the diagnostic ability of the model.

[0096] The optimization goal of the loss function in machine learning is to minimize it. The smaller the loss function, the stronger the DRN model's anti-disturbance ability. In the present invention, the Softmax loss function is selected for DRN model training. The Softmax loss function is the cross entropy loss function of Softmax. The function of Softmax is a classifier. Given an input prediction value x i And the corresponding true label value y i , and its calculation formula is:

[0097]

[0098] Among them, x i represents the input feature of the i-th one; y i Indicates the label of the category corresponding to the i-th input, Represents the bias corresponding to the true category of the i-th sample, W j represents the weight vector of the jth class, represents the weight vector of the input corresponding to the category label, T represents the vector transpose, b j Represents the bias of the sample category, and n represents the number of sample categories.

[0099] The calculation formula of Softmax loss is:

[0100]

[0101] Among them, L s represents the Softmax loss function, that is, the sample average cross entropy; m represents the number of training samples.

[0102] Step 5: Use the trained deep residual network to identify the status of power transmission and transformation equipment.

[0103] The collected time domain features of the status parameters of each power transmission and transformation equipment to be identified are input into the trained DRN model, and the deep residual network is used to achieve differentiated and refined identification of the operating status and fault type of the power transmission and transformation equipment.

[0104] The beneficial effect of the present invention is that, compared with the prior art, the deep residual network model is trained and verified based on the historical data of the operating status of the power transmission and transformation equipment, which can analyze the operating parameters of the power transmission and transformation equipment, accurately evaluate its operating status, realize accurate identification of the status of the power transmission and transformation equipment, and provide support for dispatching personnel to grasp the information of the power transmission and transformation equipment.

[0105] The applicant of the present invention has made a detailed explanation and description of the implementation examples of the present invention in conjunction with the drawings in the specification. However, those skilled in the art should understand that the above implementation examples are only preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, but not to limit the scope of protection of the present invention. On the contrary, any improvements or modifications based on the inventive spirit of the present invention should fall within the scope of protection of the present invention.

Claims

1. A device fault tolerance diagnosis method based on deep residual network, characterized in that: The method comprises the following steps: Step 1: Collect the operating status data of power transmission and transformation equipment; Step 2: Extract the feature values ​​of the collected status data at a certain time interval to obtain the time domain characteristics of each device under different operating conditions; compare the importance of different types of status quantities in the data set, delete irrelevant attributes and redundant attributes, and extract key status parameters; after selecting the key parameters, obtain the time domain characteristics of each device under different operating conditions according to the time domain analysis method; Step 3: Classify and encode the extracted time domain feature data according to the equipment type and operating status; Step 4: Construct training sets and test sets, and input the classified and encoded data into the deep residual network to train the deep network model; Step 5: Use the trained deep residual network model to differentiate and refine the status of power transmission and transformation equipment; In step 3, the extracted time domain feature data is classified according to the device type and operating status, and different types of data are encoded and labeled accordingly; In step 4, 80% of the time domain feature data of different operating states of each device are randomly selected as training sets and validation sets to perform parameter training on the deep residual network model; the remaining 20% ​​are used as test sets to analyze and test the performance of the deep residual network model; In step 4, the deep residual network includes a residual module, and multiple residual modules are connected end to end to form a deep residual network; In step 4, the residual module consists of a convolutional layer, a batch normalization layer, and an activation function layer; The convolution layer uses multiple different convolution kernels to extract the main features of the input data. For one-dimensional input data, the output of the lth convolution layer is: in, is the output of the lth convolutional layer, is the input of the lth convolutional layer, is the convolution kernel weight, is the corresponding bias; Batch normalization makes the output of each layer conform to the Gaussian distribution with the same mean and variance. For a small batch of data x i , the operation of batch normalization layer is expressed as: Among them, μ is the mean of a set of data, ε is the minimum value, and σ is the variance of the set of data. To batch normalize the data after one layer of processing; The activation function layer uses the Relu function, which is expressed as: Relu(x)=max(0,x); The final output of the residual module is: y=x+F(x,w c ) Where x is the residual block input, F is a nonlinear function including batch normalization layer, activation function layer and convolution layer, and w c The parameters that need to be optimized in the nonlinear functions including batch normalization layers, activation function layers, and convolutional layers; The deep residual network includes 9 convolutional layers, wherein the first convolutional layer adopts a wide convolutional kernel, and the remaining convolutional layers are composed of 4 residual modules, each of which contains 2 convolutional layers; A global average pooling layer is used after the residual module, and a Dropout layer is added after the global average pooling layer. Finally, the fault classification is achieved through the Softmax classifier. Equipment types include power transformers, high-voltage circuit breakers, disconnectors, voltage / current transformers, and transmission lines; The operating status includes normal status and different fault status; the fault status of power transformer includes overheating fault, winding failure, and discharge fault; the fault status of high-voltage circuit breaker includes mechanical fault, secondary circuit fault, and sealing failure fault, etc.; the fault status of disconnector includes overheating fault, mechanical fault, and insulation fault; the fault status of voltage / current transformer includes mechanical fault, insulation fault, secondary circuit fault, and resonance fault, and the fault status of transmission line includes grounding fault, circuit breaker fault, open circuit fault, and overload fault.

2. The device fault tolerance diagnosis method of the deep residual network according to claim 1, characterized in that: In step 1, the status data comes from equipment data, monitoring data and external data; Equipment data includes nameplate parameters, ledger data, operation data, inspection data, defect / elimination data, and fault / maintenance records; The detection data includes live detection data and online monitoring data; live monitoring data includes infrared detection, ultraviolet detection, and ultrasonic detection; online monitoring data includes dissolved gas in oil, partial discharge, core grounding current, and gas pressure; External data includes meteorology, environment, geography, power grid dispatch, and load.

3. The device fault tolerance diagnosis method of the deep residual network according to claim 1, characterized in that: In step 4, Step 4.1, divide the extracted time domain feature data into training set, validation set and test set according to the ratio of 6:2:2; Step 4.2: Input the training set and validation set into the constructed deep residual network in batches, and calculate the error between the predicted value and the actual result; Step 4.3: After each iteration of the training set, the validation set is used for evaluation. Step 4.4, update the weights of each layer of the DRN network according to the average cross entropy algorithm, and repeat steps 4.2 to 4.4 until the loss function is minimized to obtain the optimal deep residual network; Finally, the test set is input into the trained deep residual network, and the diagnosis results are output to verify the diagnostic ability of the model.

4. The device fault tolerance diagnosis method of the deep residual network according to claim 3, characterized in that: The loss function uses the Softmax loss function. The function of Softmax is a classifier. Given an input prediction value x i And the corresponding true label value y i , and its calculation formula is: Among them, x i represents the input feature of the i-th one; y i Indicates the label of the category corresponding to the i-th input, Represents the bias corresponding to the true category of the i-th sample, W j represents the weight vector of the jth class, represents the weight vector of the input corresponding to the category label, T represents the vector transpose, b j Represents the bias of the sample category, and n represents the number of sample categories; The calculation formula of Softmax loss is: Among them, Ls represents the Softmax loss function, that is, the sample average cross entropy; m represents the number of training samples.

Citation Information

Cited By

  • Network technology service fault diagnosis and early warning system based on multi-source data fusion

    CN122226584A