WO2024138995A1 - Electrical device state sound identification method considering fusion of time-domain and frequency-domain features - Google Patents

Abstract

An electrical device state sound identification method considering fusion of time-domain and frequency-domain features, comprising the following steps: acquiring and preprocessing sound time-domain data; transforming a time-domain signal into the frequency domain by means of Fourier transform, simultaneously performing, on the basis of an embedded method, frequency-domain feature selection and algorithm training, and selecting an optimal feature subset; performing time-domain signal feature extraction on the sound original signal by using a wrapper method; filtering and fusing the time-domain features and the frequency-domain features by means of a mutual information method; and establishing a neural network, and, on the basis of the time-domain and frequency-domain fusion signal, identifying a sound state. The present invention reserves important feature information of the time domain and frequency domain and, by means of the embedded method and the wrapper method, reduces the data dimension, and also achieves effective identification of a normal operation state, a normal gas relief state and an abnormal gas leakage state by means of a machine learning algorithm, thereby improving the identification accuracy.

Description

A method for electrical equipment status sound recognition considering time-frequency domain feature fusion Technical Field

The present invention relates to the technical field of electrical equipment state sound recognition, and in particular to an electrical equipment state sound recognition method considering time-frequency domain feature fusion.

Background technique

Power plant equipment is often large and densely distributed. In the past, relying on inspectors to discover defects could not guarantee full-cycle and full-range monitoring. By deploying sensors and machine learning recognition algorithms, minor faults and defects can be discovered immediately to prevent further development that could cause serious production accidents and economic losses.

The current equipment sound recognition methods mainly focus on main electrical equipment such as transformers, cables, and circuit breakers, and lack research on sound recognition of power plant sealing equipment. The commonly used classification and recognition method currently uses neural network recognition. Deng Aidong and others from Southeast University proposed a wind turbine bearing fault diagnosis method based on time-frequency domain convolutional networks and deep forests to obtain fault data of vibration, working conditions, speed, and load, and perform feature extraction based on the time-frequency domain convolutional network. The fault diagnosis is completed through a two-layer deep forest model (Deng Aidong, Liu Dongchuan, Yang Hongqiang, Fan Yongsheng. Wind turbine bearing fault diagnosis method based on time-frequency domain convolutional networks and deep forests [P]. Jiangsu Province: CN114964780A, 2022-08-30.); Fu Yunxiao and others from Beijing Jiaotong University proposed a rolling bearing fault diagnosis method based on multi-dimensional vibration feature fusion in the time-frequency domain. First, the vibration signal is subjected to wavelet denoising, and feature extraction is used to obtain time domain feature parameters. The energy matrix is obtained by wavelet packet decomposition and energy moment calculation, and the synthesis is mostly a feature matrix. According to The bearing state is judged by the distance between the indexes (Fu Yunxiao, Jia Limin, Lv Jinsong, Ji Changxu, Yao Dechen, Li Qian, Lu Yong. A rolling bearing fault diagnosis method based on the fusion of multi-dimensional vibration features in the time-frequency domain [P]. Beijing: CN104655423A, 2015-05-27.); Tan Fenglei et al. from the Maintenance Branch of Jiangsu Electric Power Co., Ltd. proposed a transformer latent defect diagnosis method based on sound monitoring. First, the installation position of the sound sensor is determined according to the noise attenuation law, and the transformer is judged whether there is a latent defect based on the characteristic frequency and defect evaluation index (Tan Fenglei, Zhu Chao, Chen Hao, Deng Kai, Gao Shiyu, Gong Chenlong. A transformer latent defect diagnosis method based on sound monitoring [P]. Jiangsu Province: CN113253156A, 2021-08-13.); Wu Xiaowen et al. from Hunan University of Science and Technology proposed a method using sound characteristics to diagnose latent defects of transformers. A method and system for diagnosing transformer faults using sound feature coding, and judging transformer fault results based on sound feature coding rules and combinations (Wu Xiaowen, Lu Ming, Chen Chaoyang, He Kui, Xie Bin, Tan Zhuangxi, Zou Ying, Cao Hao. A method and system for diagnosing transformer faults using sound feature coding [P]. Hunan Province: CN114527410A, 2022-05-24.); Yang Wenqiang and others from Shandong Hedui Intelligent Technology Co., Ltd. proposed a digital evaluation method and system for high-voltage bushings based on time-frequency domain feature fusion, using time domain or frequency domain evaluation units with different sampling intervals for analysis. The results are neutralized and the state of the high-voltage bushing is judged (Yang Wenqiang, Chen Xin, Zhao Fei, Liu Peng, Feng Xu. Digital evaluation method and system for high-voltage bushings based on time-frequency domain feature fusion [P]. Shandong Province: CN115015684A, 2022-09-06.); The above-mentioned invention methods do not involve the monitoring of the leakage status of the equipment. The method of the present invention proposes that the extraction of time-frequency domain signals first undergoes feature selection by the embedding method/packaging method, and finally uses the mutual information method to propose completely correlated variables, and the time-frequency domain features are simultaneously input into the neural network to realize the sound recognition of the electrical equipment status.

Summary of the invention

The present invention proposes a method for identifying the status sound of electrical equipment taking into account the fusion of time-frequency domain features. According to the data collection and feature extraction of the sensor, the time-frequency domain feature fusion of the equipment sound is completed. Combined with the powerful learning ability of the neural network, the normal working condition, normal air leakage working condition and abnormal air leakage working condition of the equipment can be identified at the same time, thereby improving the accuracy of sound recognition.

The innovation of the present invention lies in the use of embedding method, packaging method and mutual information entropy to screen and fuse time domain information and frequency domain information, and consider the fusion of time and frequency domains when identifying the sound status of the device. On the one hand, data dimensionality reduction and feature extraction are completed, and further, neural network is used for recognition to obtain better recognition effect.

The purpose of the present invention is achieved by at least one of the following technical solutions.

A method for identifying the state sound of electrical equipment by considering the fusion of time-frequency domain features comprises the following steps:

S1, sound time domain data acquisition and preprocessing;

S2, using Fourier transform to convert the time domain signal into frequency domain, and simultaneously perform frequency domain feature selection and algorithm training based on the embedding method to select the best feature subset;

S3, extracting time domain signal features from the original sound signal using a packaging method;

S4, mutual information method is used to filter and fuse time domain features and frequency domain features;

S5. Establish a neural network to identify the sound state based on the time-frequency domain fusion signal.

Further, in step S1, the sound time domain data is a sound time domain signal acquired by a sound sensor installed in the electrical equipment;

The preprocessing is to slice the acquired sound data according to time length;
_Ne = V/ _le

Among them, le _is the time domain signal length obtained by slicing the sound data according to the time length, _NE is the number of segments, and V is the sound data obtained by the sensor.

Furthermore, in step S2, the time domain signal is converted into the frequency domain using Fourier transform, and frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method to select the best feature subset, including the following steps:

S2.1, convert the original sound signal into the frequency domain using Fourier transform;

S2.2. Frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method.

Furthermore, in step S2.1, the original sound signal is converted into the frequency domain using Fourier transform, as follows:
F(ω)＝ _∫le (x)×e ^-j2πωxdx

Among them, F(ω) is the sound frequency domain information, the converted sound frequency domain information is composed of a frequency domain set, and the corresponding sound labels are composed of a target set.

Furthermore, in step S2.2, frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method, as follows:

Use the random forest algorithm model to train and evaluate the frequency domain set and target set, and filter the feature set based on the training effect, traversing all features each time;

First, a random forest algorithm model is established and an instance is initialized, that is, the number of evaluators in the random forest algorithm model is set to evaluate the model effect. The appropriate number of evaluators is set to strike a balance between training difficulty and model effect.

Then, the feature selection SelectFromModel is instantiated, that is, the random forest algorithm model initialized in the previous step is input and the hyperparameter evaluation threshold is set, and then the frequency domain set and target set are input to start the iteration. Solve on behalf of others;

According to the weight sorting of features, the feature set that meets the threshold is retained as the subsequent frequency domain feature set.

Furthermore, in step S3, the time domain signal feature extraction is performed on the original sound signal using the packaging method:

First, the random forest algorithm is initialized and the number of evaluators is set;

Instantiate feature selection, that is, input the initialized random forest algorithm, set the objective function to recursive feature elimination, set the number of features to be retained, and set the number of features to be eliminated in each iteration; then input the time domain set and the target set for solution, where the time domain set consists of the original sound data set, and the target set consists of the corresponding sound labels;

In each iteration, the features are sorted by importance and the best features are selected until a feature set that meets the required number of features is selected as the subsequent time domain feature set.

Furthermore, in step S4, the mutual information method is used to filter and fuse the time domain features and the frequency domain features:

Find the mutual information entropy of each feature variable:

Among them, R is the mutual information entropy, t is the time domain feature number, _NT is the time domain feature number, that is, the number of time domain feature sets, f is the frequency domain feature number, _NF is the frequency domain feature number, that is, the number of frequency domain feature sets, p(t,f) is the joint distribution of the target, p(t) is the marginal distribution of the time domain feature to the target, and p(f) is the marginal distribution of the frequency domain feature to the target.

Furthermore, if the information entropy is 1, it means that the two variables are completely correlated and one of them needs to be eliminated; if the information entropy is not 1, the fusion is performed in the manner of time domain features first and frequency domain features later.

Furthermore, in step S5, a neural network is established to identify the sound state based on the time-frequency domain fusion signal:

The neural network comprises an input layer, a first hidden layer and a second hidden layer connected in sequence;

The input of the input layer is the characteristic frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the state of the electrical equipment. The error is obtained by forward propagation, and the partial derivative and learning rate are used to reversely update the state. New weights;

The propagation between neural network layers is:
g＝wc+b

Among them, g is the number of neurons in the second hidden layer, c is the number of neurons in the first hidden layer, b is the bias term, and w is the weight.

Furthermore, the weight update formula is:

Among them, w` is the updated weight, L is the neural network error, and α is the learning rate.

The present invention proposes a method for electrical equipment status sound recognition that considers the fusion of time-frequency domain features, which effectively solves the problems of insufficient practicability and low reliability of current sound monitoring technology through gas monitoring and physical detection. It uses a feature screening method to complete the fusion of time-frequency features, and uses a neural network with time-frequency domain information as input to achieve accurate recognition effects.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flowchart of a method for identifying electrical equipment status sound by considering the fusion of time-frequency domain features in an embodiment of the present invention;

FIG2 is a flowchart of an algorithm step for frequency domain feature selection based on an embedding method in an embodiment of the present invention;

FIG3 is a flow chart of algorithm steps for extracting time domain signal features from original sound signals using a packaging method according to an embodiment of the present invention.

Detailed ways

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.

Embodiment 1:

A method for electrical equipment status sound recognition considering the fusion of time-frequency domain features is shown in Figure 1. Follow these steps:

S1, sound time domain data acquisition and preprocessing;

The sound time domain data is a sound time domain signal obtained by a sound sensor installed in the electrical equipment;

The preprocessing is to slice the acquired sound data according to time length;
_Ne = V/ _le

Among them, le _is the time domain signal length obtained by slicing the sound data according to the time length, _NE is the number of segments, and V is the sound data obtained by the sensor.

S2. Use Fourier transform to convert the time domain signal into frequency domain, perform frequency domain feature selection and algorithm training based on embedding method, and select the best feature subset, including the following steps:

S2.1. Use Fourier transform to convert the original sound signal into frequency domain:
F(ω)＝ _∫le (x)×e ^-j2πωxdx

Among them, F(ω) is the sound frequency domain information;

Then the converted sound frequency domain information is combined into a frequency domain set, and the corresponding sound labels are combined into a target set

S2.2. Simultaneous frequency domain feature selection and algorithm training based on embedding method:

Use the random forest algorithm model to train and evaluate the frequency domain set and target set, and filter the feature set based on the training effect, traversing all features each time;

First, a random forest algorithm model is established and an instance is initialized, that is, the number of evaluators in the random forest algorithm model is set to evaluate the model effect. The appropriate number of evaluators is set to strike a balance between training difficulty and model effect.

Then, the feature selection SelectFromModel is instantiated, that is, the random forest algorithm model initialized in the previous step is input and the hyperparameter evaluation threshold is set, and then the frequency domain set and target set are input to start iterative solution;

According to the weight sorting of features, the feature set that meets the threshold is retained as the subsequent frequency domain feature set. The algorithm flow is shown in Figure 2.

S3. Use the packaging method to extract the time domain signal features of the original sound signal:

First, the random forest algorithm is initialized and the number of evaluators is set;

Instantiate feature selection, that is, input the initialized random forest algorithm, set the objective function to recursive feature elimination, set the number of features to be retained, and set the number of features to be eliminated in each iteration; then input the time domain set and the target set for solution, where the time domain set consists of the original sound data set, and the target set consists of the corresponding sound labels;

In each iteration, the features are sorted by importance and the best features are selected until a feature set that meets the number of retained features is selected as the subsequent time domain feature set. The algorithm flow is shown in Figure 3.

S4. Use mutual information method to filter and fuse time domain features and frequency domain features:

Find the mutual information entropy of each feature variable:

Among them, R is the mutual information entropy, t is the time domain feature number, _NT is the time domain feature number, that is, the number of time domain feature sets, f is the frequency domain feature number, _NF is the frequency domain feature number, that is, the number of frequency domain feature sets, p(t,f) is the joint distribution of the target, p(t) is the marginal distribution of the time domain feature to the target, and p(f) is the marginal distribution of the frequency domain feature to the target; if the information entropy is 1, it means that the two variables are completely correlated and one of them needs to be eliminated; if the information entropy is not 1, the fusion is performed with the time domain feature first and the frequency domain feature later.

S5, establish a neural network to identify the sound state based on the time-frequency domain fusion signal;

The neural network comprises an input layer, a first hidden layer and a second hidden layer connected in sequence;

The input of the input layer is the characteristic frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the state of the electrical device, the error is calculated by forward propagation, and the weight is updated in reverse through partial derivatives and learning rates;

The propagation between neural network layers is:
g＝wc+b

Among them, g is the number of neurons in the second hidden layer, c is the number of neurons in the first hidden layer, b is the bias term, and w is the weight;

The weight update formula is:

Among them, w` is the updated weight, L is the neural network error, and α is the learning rate.

In this embodiment, sound data of three states of electrical sealing equipment are collected, namely normal operation.wav, normal leakage.wav, and abnormal leakage.wav. After 50ms slicing, the sound time domain samples are obtained, and the data dimension is [960,9600]. The sound frequency domain samples are obtained by Fourier frequency domain conversion processing, and the data dimension is [960,96000]. The number of features retained by the packaging method is set to 100, and the embedding method threshold is set to 0.005. Based on the model evaluation of the embedding method and the packaging method and the feature screening of the mutual information method, the sample space of time-frequency domain fusion is [960,130], and the time domain feature dimension is [960,100].

The neural network input layer is set with 80 nodes, the hidden layer is set with 10 neurons, and the fully connected layer is used to output the predicted label value. After 100 training times, the overall accuracy is 98.12%. The test set results are shown in Table 1.

Table 1

Embodiment 2:

The sound data of three states of electrical sealing equipment are collected, namely normal operation.MP3, normal air leakage.MP3, and abnormal air leakage.MP3. After 50ms slicing, the sound time domain samples are obtained with a data dimension of [960,9600]. The sound frequency domain samples are processed by Fourier frequency domain conversion with a data dimension of [960,96000]. The number of features retained by the packaging method is set to 100, and the embedding method threshold is set to 0.005. Based on the model evaluation of the embedding method and the packaging method and the feature screening of the mutual information method, the sample space of time-frequency domain fusion is [960,130], and the time domain feature dimension is [960,100].

The neural network input layer is set with 80 nodes, the hidden layer is set with 10 neurons, and the fully connected layer is used Output the predicted label value. After 100 training times, the overall accuracy is 97.81%. The test set results are shown in Table 2.

Table 2

Embodiment 3:

In this embodiment, sound data of three states of electrical sealing equipment are collected, namely normal operation.wav, normal leakage.wav, and abnormal leakage.wav. After 50ms slicing, the sound time domain samples are obtained, and the data dimension is [960,9600]. The sound frequency domain samples are obtained by Fourier frequency domain conversion processing, and the data dimension is [960,96000]. The number of features retained by the packaging method is set to 130, and the embedding method threshold is set to 0.002. Based on the model evaluation of the embedding method and the packaging method and the feature screening of the mutual information method, the sample space of time-frequency domain fusion is [960,160], and the time domain feature dimension is [960,130].

The neural network input layer is set with 80 nodes, the hidden layer is set with 10 neurons, and the fully connected layer is used to output the predicted label value. After 100 training times, the overall accuracy is 99.27%. The test set results are shown in Table 3.

table 3

The above-mentioned combination of identification and tracking methods is a preferred implementation mode of the present invention, but the implementation mode of the present invention is not limited to the above-mentioned embodiments. Any other modifications, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention should be equivalent replacement methods and should be included in the protection scope of the present invention.

Claims (10)

1. A method for identifying the state sound of electrical equipment by considering the fusion of time-frequency domain features, characterized in that it comprises the following steps:

   S1, sound time domain data acquisition and preprocessing;

   S2, using Fourier transform to convert the time domain signal into frequency domain, and simultaneously perform frequency domain feature selection and algorithm training based on the embedding method to select the best feature subset;

   S3, extracting time domain signal features from the original sound signal using a packaging method;

   S4, mutual information method is used to filter and fuse time domain features and frequency domain features;

   S5. Establish a neural network to identify the sound state based on the time-frequency domain fusion signal.
2. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, it is characterized in that in step S1, the sound time domain data is a sound time domain signal obtained by a sound sensor installed in the electrical equipment;

   The preprocessing is to slice the acquired sound data according to time length;
   _Ne = V/ _le

   Among them, le _is the time domain signal length obtained by slicing the sound data according to the time length, _NE is the number of segments, and V is the sound data obtained by the sensor.
3. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, it is characterized in that in step S2, the time domain signal is converted into the frequency domain by Fourier transform, and the frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method to select the best feature subset, including the following steps:

   S2.1, convert the original sound signal into the frequency domain using Fourier transform;

   S2.2. Frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method.
4. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 3, it is characterized in that in step S2.1, the original sound signal is converted into the frequency domain by Fourier transform, specifically as follows:
   F(ω)＝ _∫le (x)×e ^-j2πωxdx

   Among them, F(ω) is the sound frequency domain information, the converted sound frequency domain information is composed of a frequency domain set, and the corresponding sound labels are composed of a target set.
5. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 4, it is characterized in that in step S2.2, frequency domain feature selection and algorithm training are simultaneously performed based on the embedding method, specifically as follows:

   Use the random forest algorithm model to train and evaluate the frequency domain set and target set, and filter the feature set based on the training effect, traversing all features each time;

   First, a random forest algorithm model is established and an instance is initialized, that is, the number of evaluators in the random forest algorithm model is set to evaluate the model effect;

   Then, the feature selection SelectFromModel is instantiated, that is, the random forest algorithm model initialized in the previous step is input and the hyperparameter evaluation threshold is set, and then the frequency domain set and target set are input to start iterative solution;

   According to the weight sorting of features, the feature set that meets the threshold is retained as the subsequent frequency domain feature set.
6. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, it is characterized in that in step S3, the time domain signal feature is extracted from the original sound signal using a packaging method:

   First, the random forest algorithm is initialized and the number of evaluators is set;

   Instantiate feature selection, that is, input the initialized random forest algorithm, set the objective function to recursive feature elimination, set the number of features to be retained, and set the number of features to be eliminated in each iteration; then input the time domain set and the target set for solution, where the time domain set consists of the original sound data set, and the target set consists of the corresponding sound labels;

   In each iteration, the features are sorted by importance and the best features are selected until a feature set that meets the required number of features is selected as the subsequent time domain feature set.
7. According to the method for electrical equipment status sound recognition considering the fusion of time-frequency domain features according to claim 1, it is characterized in that in step S4, the time-domain features and the frequency-domain features are filtered and fused using the mutual information method:

   Find the mutual information entropy of each feature variable:
   

   Among them, R is the mutual information entropy, t is the time domain feature number, _NT is the time domain feature number, that is, the number of time domain feature sets, f is the frequency domain feature number, _NF is the frequency domain feature number, that is, the number of frequency domain feature sets, p(t,f) is the joint distribution of the target, p(t) is the marginal distribution of the time domain feature to the target, and p(f) is the marginal distribution of the frequency domain feature to the target.
8. According to the method for electrical equipment status sound recognition considering the fusion of time and frequency domain features as described in claim 7, it is characterized in that if the information entropy is 1, it means that the two variables are completely correlated and one of them needs to be eliminated; if the information entropy is not 1, the fusion is performed in the form of time domain features first and frequency domain features later.
9. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, it is characterized in that in step S5, a neural network is established to recognize the sound status based on the time-frequency domain fusion signal:

   The neural network comprises an input layer, a first hidden layer and a second hidden layer connected in sequence;

   The input of the input layer is the characteristic frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the state of the electrical device, the error is calculated by forward propagation, and the weight is updated in reverse through partial derivatives and learning rates;

   The propagation between neural network layers is:
   g＝wc+b

   Among them, g is the number of neurons in the second hidden layer, c is the number of neurons in the first hidden layer, b is the bias term, and w is the weight.
10. According to the method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 9, it is characterized in that the weight update formula is:
    

    Among them, w` is the updated weight, L is the neural network error, and α is the learning rate.