An electrical equipment state sound recognition method considering time-frequency domain feature fusion

Abstract

The application discloses a kind of electrical equipment state sound recognition methods considering time-frequency domain feature fusion.The method includes the following steps: sound time domain data acquisition and pre-processing;Fourier transform is used to convert time-domain signal into frequency domain, frequency domain feature selection and algorithm training are carried out simultaneously based on embedding method, and the best feature subset is selected;Time-domain signal feature extraction is carried out on the original sound signal using packing method;Mutual information method is used to filter and fuse time-domain features and frequency-domain features;A neural network is established, and the sound state is identified based on time-frequency domain fusion signal.The application effectively fuses time-domain information and frequency-domain information, on the one hand, retains important feature information of time domain and frequency domain, reduces data dimension through embedding method and packing method, on the other hand, effectively identifies three states of normal operation, normal air leakage and abnormal air leakage, and improves the accuracy of identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment state sound recognition, and particularly relates to an electrical equipment state sound recognition method considering time-frequency domain feature fusion. BACKGROUND

[0002] Power plant equipment is often large in size and densely distributed. In the past, it was impossible to ensure full-cycle and full-range monitoring by relying on the patrol of inspectors to find defects. By laying sensors and machine learning recognition algorithms, small faults and defects can be found, and further development can be prevented to cause serious production accidents and economic losses.

[0003] The current device sound recognition method mainly studies the transformer, cable, circuit breaker and other electrical main equipment, and lacks sound recognition research on sealed equipment in power plants. The commonly used classification recognition method adopts neural network recognition. Deng Aidong and others from Southeast University proposed a wind power bearing fault diagnosis method based on time-frequency domain convolution network and deep forest, obtained fault data of vibration, working condition, rotating speed and load, extracted features according to the time-frequency domain convolution network, and completed fault diagnosis through a two-layer deep forest model (Deng Aidong, Liudongchuan, Yang Hongqiang, Fan Yongsheng. Wind power bearing fault diagnosis method based on time-frequency domain convolution network and deep forest [P]. Jiangsu province: CN114964780A, 2022-08-30.); Fu Yunxiao and others from Beijing Jiaotong University proposed a rolling bearing fault diagnosis method based on time-frequency domain multi-dimensional vibration feature fusion, first carried out wavelet denoising processing on the vibration signal, obtained time domain feature parameters by using feature extraction, obtained energy matrix by using wavelet packet decomposition and energy moment calculation, synthesized multi-dimensional feature matrix, and judged the bearing state according to the index distance (Fu Yunxiao, Jia Limin, Lv Jinsong, Ji Changxu, Yao Dechen, Li Qian, Lu Yong. Rolling bearing fault diagnosis method based on time-frequency domain multi-dimensional vibration feature fusion [P]. Beijing: CN104655423A, 2015-05-27.); Tan Fenglei and others from the Maintenance Branch of Jiangsu Electric Power Co., Ltd. proposed a transformer latent defect diagnosis method based on sound monitoring, first determined the installation position of the sound sensor according to the noise attenuation law, and judged whether the transformer had latent defects based on the characteristic frequency and defect evaluation index (Tan Fenglei, Zhu Chao, Chen Hao, Deng Kai, Gao Shiyu, Gong Chenlong. Transformer latent defect diagnosis method based on sound monitoring [P]. Jiangsu province: CN113253156A, 2021-08-13.); Wu Xiaowen and others from Hunan University of Science and Technology proposed a method and system for diagnosing transformer faults by using sound feature coding, and judged the transformer fault result according to the sound feature coding rule and combination (Wu Xiaowen, Lu Ming, Chen Chaoyang, He Qi, Xie Bin, Tan Zhuangxi, Zou Ying, Cao Hao. Method and system for diagnosing transformer faults by using sound feature coding [P]. Hunan province: CN114527410A, 2022-05-24.); Yang Wenqiang and others from Shandong Hechong Intelligence Technology Co., Ltd. proposed a high-voltage sleeve digital evaluation method and system based on time-frequency domain feature fusion, and analyzed the results of the time domain or frequency domain evaluation unit in different sampling intervals to judge the state of the high-voltage sleeve (Yang Wenqiang, Chen Xin, Zhao Fei, Liu Peng, Feng Xu. High-voltage sleeve digital evaluation method and system based on time-frequency domain feature fusion [P]. Shandong province: CN115015684A, 2022-09-06.); The above invention methods do not involve monitoring the air leakage state of the equipment. SUMMARY

[0004] The application provides an electrical equipment state sound recognition method considering time-frequency domain feature fusion, which realizes time-frequency domain feature fusion of equipment sound according to sensor data acquisition and feature extraction, combines the powerful learning ability of a neural network, can simultaneously recognize normal working condition, normal air leakage working condition and abnormal air leakage working condition of the equipment, and improves the accuracy of sound recognition.

[0005] The application has the advantages that the time domain information and the frequency domain information are screened and fused by using embedding method, packaging method and mutual information entropy, the time-frequency domain fusion is considered in recognizing the equipment sound state, data dimension reduction and feature extraction are realized, and better recognition effect is obtained by using the neural network for recognition.

[0006] The object of the application is achieved at least by one of the following technical solutions.

[0007] An electrical equipment state sound recognition method considering time-frequency domain feature fusion comprises the following steps.

[0008] S1, sound time domain data acquisition and preprocessing;

[0009] S2, converting the time domain signal into a frequency domain by using Fourier transform, simultaneously performing frequency domain feature selection and algorithm training based on embedding method, and selecting a best feature subset;

[0010] S3, extracting time domain signal features from the original sound signal by using packaging method;

[0011] S4, filtering and fusing the time domain features and the frequency domain features by using mutual information method;

[0012] S5, establishing a neural network and recognizing the sound state based on the time-frequency domain fusion signal.

[0013] Further, in the step S1, the sound time domain data is a sound time domain signal acquired from a sound sensor arranged on the electrical equipment;

[0014] The preprocessing is to slice the acquired sound data according to time length;

[0015] N e = V / l e

[0016] Wherein, l e is the time domain signal length obtained by slicing the sound data according to time length, N E is the number of segments, and V is the sound data acquired by the sensor.

[0017] Further, in the step S2, the time domain signal is converted into a frequency domain by using Fourier transform, the frequency domain feature selection and algorithm training are simultaneously performed based on embedding method, and the best feature subset is selected, comprising the following steps:

[0018] S2.1, converting the original sound signal to the frequency domain by using Fourier transform;

[0019] S2.2, simultaneously performing frequency domain feature selection and algorithm training based on embedding method.

[0020] Further, in step S2.1, the original sound signal is converted to the frequency domain by using Fourier transform, specifically as follows:

[0021] F(ω) = ∫l e (x) × e -j2πωx dx

[0022] Wherein, F(ω) is the sound frequency domain information, the converted sound frequency domain information is composed of the frequency domain set, and the corresponding sound label is composed of the target set.

[0023] Further, in step S2.2, the frequency domain feature selection and algorithm training are simultaneously performed based on the embedding method, specifically as follows:

[0024] The random forest algorithm model is used to train and evaluate the effect of the frequency domain set and the target set, and the feature set is filtered according to the training effect, and all features are searched and traversed each time;

[0025] First, the random forest algorithm model is established and instance initialization is performed, that is, the number of evaluators in the random forest algorithm model is set to evaluate the model effect, and a proper number of evaluators is set to balance the training difficulty and the model effect;

[0026] Then, the instantiation of feature selection SelectFromModel is performed, that is, the random forest algorithm model initialized in the last step and the evaluation threshold of hyperparameter are input, and then the frequency domain set and the target set are input to start iteration solution;

[0027] According to the weight order of the features, the feature set meeting the threshold is reserved as the subsequent frequency domain feature set.

[0028] Further, in step S3, the original sound signal is processed by using the wrapping method to extract the time domain signal features:

[0029] First, the random forest algorithm is instance initialized, and the number of evaluators is set;

[0030] The instantiation of feature selection is performed, that is, the initialized random forest algorithm is input, the objective function is set to recursive feature elimination method, the number of reserved features is set, and the number of features to be removed in each iteration is set; Then input the time domain set and the target set for solution, wherein the time domain set is composed of the original sound data set, and the target set is composed of the corresponding sound label;

[0031] In each iteration, features are sorted by importance, and the best features are selected until a feature set that meets the requirement of retaining the required number of features is selected as the subsequent time-domain feature set.

[0032] Furthermore, in step S4, the mutual information method is used to filter and fuse the time-domain features and frequency-domain features:

[0033] Calculate the mutual information entropy for each feature variable:

[0034]

[0035] Where R is the mutual information entropy, t is the time-domain feature index, and N is the number of features. T N is the number of time-domain features, i.e., the number of time-domain feature sets, f is the frequency-domain feature index, and N is the number of frequency-domain features. F denoted as the number of frequency domain features, i.e., 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 features to the target, and p(f) is the marginal distribution of the frequency domain features to the target.

[0036] Furthermore, if the information entropy is 1, it means that the two variables are completely correlated and one of them needs to be removed; if the information entropy is not 1, then the fusion is performed in the manner of time domain features first and frequency domain features second.

[0037] Furthermore, in step S5, a neural network is established to identify the sound state based on the time-frequency domain fusion signal:

[0038] The neural network includes an input layer, a first hidden layer, and a second hidden layer connected in sequence;

[0039] The input to the input layer is the feature frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the electrical equipment status. The error is calculated using forward propagation, and the weights are updated in reverse using partial derivatives and learning rate.

[0040] Inter-layer propagation in a neural network is as follows:

[0041] g = wc + b

[0042] Where 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.

[0043] Furthermore, the weight update formula is as follows:

[0044]

[0045] Where w' is the updated weight, L is the neural network error, and α is the learning rate.

[0046] Compared with the prior art, the advantages of this invention are:

[0047] The application provides an electrical equipment state sound recognition method considering time-frequency domain feature fusion, effectively solves the problems of insufficient practicability and low reliability of current sound monitoring technology through gas monitoring and physical detection, uses a feature screening method to complete time-frequency feature fusion, and uses a neural network taking time-frequency domain information as input to achieve accurate recognition effect. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A step flowchart of the electrical equipment state sound recognition method considering time-frequency domain feature fusion in the embodiment of the application is shown in the figure.

[0049] Figure 2 An algorithm step flowchart of the frequency domain feature selection based on the embedding method in the embodiment of the application is shown in the figure.

[0050] Figure 3 An algorithm step flowchart of the time domain signal feature extraction of the sound original signal by the packing method in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the application will be clearly and completely described in combination with the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0052] Embodiment 1:

[0053] An electrical equipment state sound recognition method considering time-frequency domain feature fusion, as shown in the figure, includes the following steps: Figure 1

[0054] S1, sound time domain data acquisition and preprocessing;

[0055] The sound time domain data is a sound time domain signal acquired according to a sound sensor installed on the electrical equipment;

[0056] The preprocessing is time length slicing on the acquired sound data;

[0057] N e = V / l e

[0058] Wherein, l e is the time domain signal length obtained by time length slicing on the sound data, N E is the number of segments, and V is the sound data acquired by the sensor.

[0059] ​S2, convert the time domain signal into frequency domain by Fourier transform, simultaneously perform frequency domain feature selection and algorithm training based on embedding method, select the best feature subset, including the following steps:

[0060] S2.1, convert the original sound signal into frequency domain by Fourier transform:

[0061] F(ω) = ∫l e (x) x e -j2πωx dx

[0062] Wherein, F(ω) is the sound frequency domain information;

[0063] Then the converted sound frequency domain information is composed of a frequency domain set, and the corresponding sound label is composed of a target set

[0064] S2.2, simultaneously perform frequency domain feature selection and algorithm training based on embedding method:

[0065] Use random forest algorithm model to train and evaluate the effect of frequency domain set and target set, and filter the feature set according to the training effect, search all features each time;

[0066] First, establish a random forest algorithm model and perform instance initialization, that is, set the number of evaluators in the random forest algorithm model for evaluating the model effect, and set a reasonable number of evaluators to balance the training difficulty and model effect;

[0067] Then, the instantiation of feature selection SelectFromModel is performed, that is, the random forest algorithm model initialized in the last step and the evaluation threshold value of the super parameter are input, and then the frequency domain set and the target set are input to start iteration solution;

[0068] According to the weight order of the features, the feature set meeting the threshold value is reserved as the subsequent frequency domain feature set, and the algorithm flow is shown in Figure 2 .

[0069] S3, use wrapping method to extract time domain signal features of sound original signal:

[0070] First, instance initialization of random forest algorithm, set the number of evaluators;

[0071] The instantiation of feature selection is performed, that is, the initialized random forest algorithm is input, the objective function is set to recursive feature elimination method, the number of reserved features is set, and the number of features to be eliminated in each iteration is set; Then input the time domain set and the target set for solution, wherein the time domain set is composed of the original sound data set, and the target set is composed of the corresponding sound label;

[0072] In each iteration, the best features are selected according to the feature importance until a feature set meeting the reserved feature number is selected as the subsequent time domain feature set, and the algorithm flow is shown in Figure 3 .

[0073] S4, mutual information method is used to filter and fuse the time domain features and the frequency domain features:

[0074] The mutual information entropy of each feature variable is calculated:

[0075]

[0076] Wherein, R is the mutual information entropy, t is the time domain feature serial number, N T is the time domain feature number, i.e., the number of the time domain feature set, f is the frequency domain feature serial number, N F is the frequency domain feature number, i.e., the number of the frequency domain feature set, p(t,f) is the joint distribution of the target, p(t) is the edge distribution of the time domain feature to the target, and p(f) is the edge distribution of the frequency domain feature to the target; if the information entropy is 1, it means that the two variables are completely related, and one of them needs to be removed; if the information entropy is not 1, the fusion is performed in the order of the time domain feature first and the frequency domain feature second.

[0077] S5, a neural network is established, and the sound state is recognized based on the time-frequency domain fused signal;

[0078] The neural network comprises an input layer, a first hidden layer and a second hidden layer connected in sequence;

[0079] The input of the input layer is the feature frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the electrical equipment state, the error is calculated by forward propagation, and the weight is updated in reverse through the partial derivative and the learning rate;

[0080] The neural network layer propagation is:

[0081] g=wc+b

[0082] Wherein, g is the number of neurons of the second hidden layer, c is the number of neurons of the first hidden layer, b is the bias term, and w is the weight;

[0083] The weight update formula is:

[0084]

[0085] Wherein, w' is the updated weight, L is the neural network error, and a is the learning rate.

[0086] In this embodiment, the sound data of three states of the electrically sealed device is collected, which are normal working.wav, normal air leakage.wav and abnormal air leakage.wav. After 50 ms 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 reserved features of the packaging method is set to 100, and the threshold of the embedding method is set to 0.005. Based on the model evaluation and mutual information method feature screening of the embedding method and the packaging method, the time-frequency domain fusion sample space is [960, 130], wherein the time domain feature dimension is [960, 100].

[0087] The input layer of the neural network is set to 80 nodes, the hidden layer is set to 10 neurons, the predicted label value is output by the full connection layer, and after 100 times of training, the overall accuracy is 98.12%, and the test set results are shown in Table 1.

[0088] Table 1

[0089] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 98.3％ 99.1％ 97.3％

[0090] Embodiment 2:

[0091] The sound data of three states of the electrically sealed device is collected, which are normal working.MP3, normal air leakage.MP3 and abnormal air leakage.MP3. After 50 ms 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 reserved features of the packaging method is set to 100, and the threshold of the embedding method is set to 0.005. Based on the model evaluation and mutual information method feature screening of the embedding method and the packaging method, the time-frequency domain fusion sample space is [960, 130], wherein the time domain feature dimension is [960, 100].

[0092] The input layer of the neural network is set to 80 nodes, the hidden layer is set to 10 neurons, the predicted label value is output by the full connection layer, and after 100 times of training, the overall accuracy is 97.81%, and the test set results are shown in Table 2.

[0093] Table 2

[0094] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 98.1％ 99.7％ 95.3％

[0095] Embodiment 3:

[0096] In this embodiment, the sound data of three states of the electrically sealed device 3 is collected, which are normal working.wav, normal air leakage.wav and abnormal air leakage.wav. After 50 ms slicing, the sound time domain samples are obtained, and the data dimension is [960, 9600]. The Fourier frequency domain conversion processing is used to obtain the sound frequency domain samples, and the data dimension is [960, 96000]. The number of reserved features of the packaging method is set to 130, and the threshold of the embedding method is set to 0.002. Based on the model evaluation and mutual information method feature screening of the embedding method and the packaging method, the time-frequency domain fusion sample space is [960, 160], wherein the time domain feature dimension is [960, 130].

[0097] The neural network input layer is set to 80 nodes, the hidden layer is set to 10 neurons, the full connection layer is used to output the predicted label value, and after 100 times of training, the overall accuracy is 99.27%, and the test set results are shown in Table 3.

[0098] Table 3

[0099] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 99.1％ 99.0％ 99.1％

[0100] The above-mentioned identification and tracking method is combined as the preferred embodiment of the present application, but the embodiments of the present application are not limited by the above-mentioned embodiments, and any modification, modification, substitution, combination, simplification made without departing from the spirit and principles of the present application should be an equivalent replacement mode, and should be included in the protection scope of the present application.

Claims

1. A method for recognizing the status sound of electrical equipment considering time-frequency domain feature fusion, characterized in that, Includes the following steps: S1. Acquisition and preprocessing of audio time-domain data; S2. The time-domain signal is converted into the frequency domain using Fourier transform. Frequency domain feature selection and algorithm training are performed simultaneously based on the embedding method to select the best feature subset. The time-domain signal is converted into the frequency domain using Fourier transform. Based on the embedding method, frequency domain feature selection and algorithm training are performed simultaneously to select the optimal feature subset. The steps include: S2.

1. Use Fourier transform to convert the original sound signal to the frequency domain; the specific steps are as follows: Wherein, F(ω) represents the audio frequency domain information, the converted audio frequency domain information is used to form a frequency domain set, and the corresponding audio tags are used to form a target set; S2.2 Simultaneous Frequency Domain Feature Selection and Algorithm Training Based on Embedding Method; The specific details are as follows: The random forest algorithm model is used to train and evaluate the frequency domain set and the target set, and the feature set is filtered according to the training effect. Each search traverses all features. First, a random forest algorithm model is established and instances are initialized, that is, the number of evaluators in the random forest algorithm model is set to evaluate the model's performance. Then, the SelectFromModel feature selection is instantiated, which means inputting the random forest algorithm model initialized in the previous step and setting the hyperparameter evaluation threshold, and then inputting the frequency domain set and the target set to start iterative solution; Based on the weights of the features, the feature set that meets the threshold is retained as the subsequent frequency domain feature set; S3. Use the packaging method to extract time-domain signal features from the original sound signal; S4. Mutual information method filters and fuses time-domain and frequency-domain features; S5. Establish a neural network to identify the sound state based on the time-frequency domain fusion signal.

2. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, characterized in that, In step S1, the sound time-domain data is the sound time-domain signal obtained from the sound sensor installed in the electrical equipment; The preprocessing involves slicing the acquired audio data by duration. Among them, l e N represents the length of the time-domain signal obtained by slicing the audio data by duration. E V represents the number of segments, and V represents the sound data acquired by the sensor.

3. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, characterized in that, In step S3, the time-domain signal features of the original sound signal are extracted using the packaging method: First, initialize the random forest algorithm instance and set the number of evaluators; Instantiate the feature selection process by inputting the initialized random forest algorithm, setting the objective function to recursive feature elimination, setting the number of features to retain, and setting the number of features to remove in each iteration; then input the time-domain set and the target set for solving, where the time-domain set consists of the original sound dataset and the target set consists of the corresponding sound tags; In each iteration, features are sorted by importance, and the best features are selected until a feature set that meets the requirement of retaining the required number of features is selected as the subsequent time-domain feature set.

4. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, characterized in that, In step S4, the mutual information method is used to filter and fuse the time-domain features and frequency-domain features: Calculate the mutual information entropy for each feature variable: Where R is the mutual information entropy, t is the time-domain feature index, and N is the number of features. T N is the number of time-domain features, i.e., the number of time-domain feature sets, f is the frequency-domain feature index, and N is the number of frequency-domain features. F denoted as the number of frequency domain features, i.e., 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 features to the target, and p(f) is the marginal distribution of the frequency domain features to the target.

5. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 4, 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 removed; if the information entropy is not 1, then the variables are fused in the manner of time domain features first and frequency domain features second.

6. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 1, characterized in that, In step S5, a neural network is established to identify the sound state based on the time-frequency domain fusion signal: The neural network includes an input layer, a first hidden layer, and a second hidden layer connected in sequence; The input to the input layer is the feature frequency band selected in step S4; the output of the first hidden layer and the second hidden layer is the electrical equipment status. The error is calculated using forward propagation, and the weights are updated in reverse using partial derivatives and learning rate. Interlayer propagation in a neural network is as follows: g = wc + b Where 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.

7. The method for electrical equipment status sound recognition considering time-frequency domain feature fusion according to claim 6, characterized in that, The weight update formula is: Where w' is the updated weight, L is the neural network error, and α is the learning rate.

Images