CN115983265A - A Convolutional Neural Network Based Defect Grading Method for Relay Protection - Google Patents

Abstract

The invention relates to the technical field of defect analysis of relay protection devices, and provides a method for grading defects of relay protection devices based on a convolutional neural network. The method includes: preprocessing the data based on the defect records of relay protection operation in a certain regional power grid, and obtaining a defect grading data set, and using a one-dimensional convolutional layer to extract features from the vectorized text matrix through the Markov assumption, Pass the obtained features into the fully connected neural network to obtain the text prediction result of recognition vectorization, and perform the preset number of model iterations to convergence through the forward propagation gradient, and obtain the target convolutional neural network model; based on the selected prediction parameters , input the test set into the target convolutional neural network model to obtain the recognition results of defect grading. The invention improves the classification accuracy rate of the grading prediction result of the defect record text.

Description

A Convolutional Neural Network Based Defect Grading Method for Relay Protection

technical field

The invention relates to the technical field of defect analysis of relay protection devices, in particular to a method for grading defects of relay protection devices based on a convolutional neural network.

Background technique

During the long-term operation of relay protection equipment, a large amount of defect text data has been recorded and accumulated through inspections, tests and other means. After the text data is stored in the system, it is usually used as record data, and a large amount of valuable information contained in it has not been discovered, which is the key data for the classification of equipment defect levels.

However, a large amount of equipment defect classification work needs to be done manually, which not only has low efficiency and heavy workload, but also is often in an embarrassing situation where it is difficult to accurately judge some sub-health defects with strong ambiguity. Therefore, the classification accuracy is limited. Influence.

Contents of the invention

In view of this, the present invention provides a method for grading defects in relay protection based on convolutional neural networks to solve the problems in the prior art that the classification of defect levels needs to be done manually, the efficiency is low, and the classification accuracy is poor.

The present invention provides a method for grading relay protection faults based on convolutional neural network, including:

S1. Preprocessing the data based on the defect records of relay protection operation in a certain area to obtain a defect grading data set, wherein the defect grading data set includes a training set, a test set and a verification set;

S2. Through the Markov assumption, a one-dimensional convolutional layer is used to extract features from the vectorized text matrix, and the obtained features are passed into the fully connected neural network to obtain the prediction result of the vectorized text recognition, and the gradient is propagated forward , performing a preset number of model iterations to convergence to obtain the target convolutional neural network model;

S3. Based on the selected prediction parameters, input the test set into the target convolutional neural network model to obtain a defect grading recognition result.

Further, the S1 includes:

S11. Based on the defect records of relay protection operation in a certain regional power grid, use the relay protection defect dictionary to remove stop words and irrelevant words, and perform vectorization processing to form a vectorized text matrix, and convert the vectorized text matrix Divided into three defect levels: critical, serious and general;

S12. Utilize the method of jieba word segmentation in combination with professional dictionaries to carry out word segmentation processing on the three defect levels of critical, serious and general respectively, and divide them into the training set, test set and verification set according to 6:2:2, wherein jieba is A word segmentation function package based on the python programming language.

Further, the model training in S2 includes: batch_size=128, and preset times=10000 times.

Further, said S2 includes:

S21. According to the Markov assumption, construct three one-dimensional convolution kernels with lengths of 2, 3, and 4;

S22. Using the constructed one-dimensional convolution kernel, perform a convolution operation on the vector text matrix, extract the feature part containing a set amount of information entropy in the text, and pass the feature part into the fully connected neural network , using sigmoid as the activation function to calculate, and then using the softmax function to output the probability distribution of the category of the vectorized text to obtain the text prediction result of the recognition vectorization;

S23. Based on the text prediction result of the recognition vectorization, use a cross-entropy loss function to propagate the gradient forward, perform a preset number of model iterations until convergence, and obtain the target convolutional neural network model.

Further, the probability distribution of the category to which the vectorized text belongs is in the range [0,1], and the sum is 1.

Further, the cross-entropy loss function includes the following calculation formula:

Among them, M represents the number of categories, y _ic represents the sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0, and p _ic represents the probability that observed sample i belongs to category c.

Further, said S3 includes:

S31. Accuracy, recall, and F1 score are used to represent the prediction accuracy of the target convolutional neural network model;

S32. Input the test set into the target convolutional neural network model to obtain a defect grading recognition result.

Further, the acquisition of the model evaluation index includes the following calculation formula:

Among them, P represents the precision rate, R represents the recall rate, F represents the F1 score, TP represents the number of flawed texts correctly classified, FP represents the number of faulty texts classified incorrectly, and FN represents the number of non-detected texts.

The beneficial effect that the present invention exists compared with prior art is:

1. The present invention uses the softmax function to calculate, and outputs the probability distribution of the category to which the vectorized text belongs, which improves the classification accuracy of the grading prediction result of the defect record text.

2. The present invention obtains the loss function converging to the preset value by extracting the characteristic part of the text containing the set amount of information entropy, the calculation speed is fast, and the situation of falling into the local optimal solution is avoided.

3. The present invention guarantees the prediction accuracy of the target convolutional neural network model by selecting a series of model evaluation index parameters such as accuracy rate, recall rate, and F1 score, and ensures that defects can be processed and reported in a timely manner.

Description of drawings

In order to illustrate the technical solution in the present invention more clearly, the accompanying drawings that need to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some implementations of the present invention For example, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a flow chart of a method for grading defects in relay protection based on convolutional neural network provided by the present invention.

Detailed ways

In the following description, specific details such as specific system structures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention. It will be apparent, however, to one skilled in the art that the invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

A convolutional neural network-based relay protection defect grading method provided by the present invention will be described in detail below with reference to the accompanying drawings.

Fig. 1 is a flow chart of a method for grading defects in relay protection based on convolutional neural network provided by the present invention.

As shown in Figure 1, the method for grading relay protection faults includes:

S1. Preprocessing the data based on the defect records of relay protection operation in a certain area to obtain a defect grading data set, wherein the defect grading data set includes a training set, a test set and a verification set;

The S1 includes:

S11. Based on the defect records of relay protection operation in a certain regional power grid, use the relay protection defect dictionary to remove stop words and irrelevant words, and perform vectorization processing to form a vectorized text matrix, and convert the vectorized text matrix Divided into three defect levels: critical, serious and general;

S12. Utilize the method of jieba word segmentation in combination with professional dictionaries to carry out word segmentation processing on the three defect levels of critical, serious and general respectively, and divide them into the training set, test set and verification set according to 6:2:2, wherein jieba is A word segmentation function package based on the python programming language.

The defect grading data set information is shown in Table 1:

data set Training set test set validation set total critical (label 0) 972 301 290 1563 severe (label 1) 705 238 188 1131 General (Label 2) 723 239 242 1204 total 2400 778 720 3898

S2. Through the Markov assumption, a one-dimensional convolutional layer is used to extract features from the vectorized text matrix, and the obtained features are passed into the fully connected neural network to obtain the prediction result of the vectorized text recognition, and the gradient is propagated forward , performing a preset number of model iterations to convergence to obtain the target convolutional neural network model;

The model training in S2 includes: batch_size=128, and preset times=10000 times.

The S2 includes:

S21. According to the Markov assumption, construct three one-dimensional convolution kernels with lengths of 2, 3, and 4;

According to the Markov assumption: the probability of any word appearing is only related to one or a few words appearing before it, and three one-dimensional convolution kernels with lengths of 2, 3, and 4 are constructed, that is, considering the 2 , 3, 4 words influence on it.

S22. Using the constructed one-dimensional convolution kernel, perform a convolution operation on the vector text matrix, extract the feature part containing a set amount of information entropy in the text, and pass the feature part into the fully connected neural network , using sigmoid as the activation function to calculate, and then using the softmax function to output the probability distribution of the category of the vectorized text to obtain the text prediction result of the recognition vectorization;

Specifically, use the constructed one-dimensional convolution kernel to perform a convolution operation on the vectorized text matrix, extract the feature part of the text that contains a set amount of information entropy, and input the extracted feature part into a 5-layer Fully connected neural network, using sigmoid as the activation function, namely

Wherein, the fully connected neural network refers to a convolutional neural network model, and the convolutional neural network model includes multiple one-dimensional convolutions to obtain N-gram feature representations in sentences.

After the fully connected neural network outputs the probability distribution of the category of the vectorized text, and then uses the softmax function to calculate, the probability is converted into a probability distribution ranging between [0,1] and 1, wherein the acquisition of the softmax function Including the following calculation formula:

Among them, Z _i is the output value of the i-th node, and C is the number of output nodes, that is, the number of classification categories.

S23. Based on the text prediction result of the recognition vectorization, use a cross-entropy loss function to propagate the gradient forward, perform a preset number of model iterations until convergence, and obtain the target convolutional neural network model.

Use the cross-entropy loss function to calculate the model gap between the predicted classification and the actual classification, where the cross-entropy loss function includes the following calculation formula:

Among them, M is the number of categories, y _ic is a sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0, and p _ic is the probability that observed sample i belongs to category c.

Based on the stochastic gradient descent method, the cross-entropy loss function is used as the objective function, samples are continuously randomly selected to calculate the gradient, and the gradient is descended according to the set step size until the cross-entropy loss function converges to the set value, and the obtained Describe the target convolutional neural network model. The present invention adopts the random extraction of the gradient calculation of the stochastic gradient descent method, is not easy to fall into a local optimal solution, and has a relatively fast calculation speed.

S3. Based on the selected prediction parameters, input the test set into the target convolutional neural network model to obtain a defect grading recognition result.

The S3 includes:

S31. Accuracy, recall, and F1 score are used to represent the prediction accuracy of the target convolutional neural network model;

S32. Input the test set into the target convolutional neural network model to obtain a defect grading recognition result.

The target convolutional neural network model reflects the identification result of defect grading through the evaluation index of the target convolutional neural network model.

The acquisition of the target convolutional neural network model evaluation index includes the following calculation formula:

Among them, P represents the precision rate, R represents the recall rate, F represents the F1 score, TP represents the number of flawed texts correctly classified, FP represents the number of faulty texts classified incorrectly, and FN represents the number of non-detected texts.

In order to comprehensively evaluate the accuracy of the training model, set batch_size=128 during the training process. After each batch is trained for one round, the optimal model is taken out as the basic model for the next round of training. After 10,000 times of training, the following results are obtained:

P=0.6833

R=0.7560

F=0.2472

The present invention uses the softmax function to calculate, and outputs the probability distribution of the category to which the vectorized text belongs, which improves the classification accuracy rate of identifying the predicted results of the vectorized text; and obtains the convergence of the loss function by extracting the characteristic part containing the set amount of information entropy in the text To the preset value, the calculation speed is fast, and the situation of falling into the local optimal solution is avoided; by selecting a series of model evaluation parameters such as accuracy rate, recall rate, and F1 score, the prediction accuracy of the target convolutional neural network model is guaranteed. It also ensures that defects can be processed and reported in a timely manner.

All the above optional technical solutions may be combined in any way to form optional embodiments of the present application, which will not be repeated here.

It should be understood that the sequence numbers of the steps in the above embodiments do not mean the order of execution, and the execution order of each process should be determined by its functions and internal logic, and should not constitute any limitation to the implementation process of the embodiment of the present invention.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be described in the foregoing embodiments Modifications to the technical solutions recorded, or equivalent replacement of some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention, and shall be included in the scope of the technical solutions of the embodiments of the present invention. within the scope of protection.

Claims (8)

1. A relay protection defect setting and grading method based on a convolutional neural network is characterized by comprising the following steps:

s1, preprocessing data based on a certain regional power grid relay protection operation defect record to obtain a defect grading data set, wherein the defect grading data set comprises a training set, a testing set and a verification set;

s2, performing feature extraction on a text matrix subjected to opposite quantization by adopting a one-dimensional convolutional layer through Markov assumption, transmitting the obtained features into a fully-connected neural network to obtain a text prediction result subjected to identification vectorization, and performing model iteration of preset times to convergence through a forward propagation gradient to obtain a target convolutional neural network model;

and S3, inputting the test set into the target convolutional neural network model based on the selected prediction parameters to obtain a defect grading identification result.

2. The relay protection device defect grading method according to claim 1, wherein S1 comprises:

s11, based on a relay protection operation defect record of a certain regional power grid, removing stop words and irrelevant words by adopting a relay protection defect dictionary, performing vectorization processing to form a vectorized text matrix, and dividing the vectorized text matrix into three defect grades of critical, serious and general;

s12, performing word segmentation processing on the critical, serious and general three defect levels respectively by using a method of combining the jieba word segmentation with a professional dictionary, and dividing the three defect levels into the training set, the testing set and the verification set according to the following ratio of 6.

3. The relay protection device defect grading method according to claim 1, wherein the model training in S2 comprises: batch _ size =128, and the preset number = 10000.

4. The relay protection device defect grading method according to claim 1, wherein the S2 comprises:

s21, constructing 3 one-dimensional convolution kernels with the length of 2,3,4 according to the Markov assumption;

s22, performing convolution operation on the vector text matrix by using the constructed one-dimensional convolution kernel, extracting a characteristic part containing a set amount of information entropy in a text, transmitting the characteristic part into the fully-connected neural network, calculating by using sigmoid as an activation function, and outputting probability distribution of the category to which the vectorized text belongs by using a softmax function to obtain a text prediction result for identifying vectorization;

and S23, based on the recognition vectorization text prediction result, performing model iteration of preset times to convergence by using a cross entropy loss function and through a forward propagation gradient to obtain the target convolutional neural network model.

5. The relay protection device defect grading method according to claim 4, wherein the probability distribution of the category to which the vectorized text belongs is in the range of [0,1] and is 1.

6. The relay protection device defect grading method according to claim 4, wherein the cross entropy loss function comprises the following calculation formula:

where M denotes the number of categories, yi _c Representing a symbolic function, taking 1 if the true class of sample i is equal to c, otherwise taking 0 _c Representing the probability that the observed sample i belongs to the class c.

7. The relay protection device defect grading method according to claim 1, wherein the S3 comprises:

s31, representing the prediction precision of the target convolutional neural network model by adopting the accuracy, the recall rate and the F1 score;

and S32, inputting the test set into the target convolutional neural network model to obtain a defect classification recognition result.

8. The relay protection device defect grading method according to claim 7, wherein the obtaining of the model evaluation index comprises the following calculation formula:

wherein P represents accuracy, R represents recall, F represents F1 score, TP represents number of correctly ranked defect texts, FP represents number of incorrectly ranked defect texts, and FN represents number of undetected texts.

Images