An electrical equipment sound recognition method based on feature frequency band selection

Abstract

The application discloses an electrical equipment sound recognition method based on feature frequency band selection. The method comprises the following steps: sound time domain data acquisition and preprocessing, conversion to a frequency domain by using Fourier transform; a Metropolis-Hastings sampling method is used to randomly sample a set number of samples to establish a partial least squares regression model; a logarithmic attenuation function is used to filter the frequency band with the minimum absolute value weight of the regression coefficient; a set retention ratio of the frequency band is selected by using weighted sampling, and the corresponding frequency band with a smaller root mean square error of cross-validation is selected as a feature frequency band; a neural network is established to recognize the sound state. The application reduces the spatial dimension of the training sample, avoids data redundancy, and improves the training speed. The application effectively recognizes three states of normal working, normal air leakage and abnormal air leakage, and improves the recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment state recognition, and particularly relates to an electrical equipment sound recognition method based on feature frequency band selection. BACKGROUND

[0002] Normal operation of the electrical main equipment, electrical auxiliary equipment and action equipment is crucial to the production of the power plant, and full-cycle and full-range monitoring of the equipment can timely find potential defects and prevent further development to cause serious production accidents and economic losses.

[0003] The current equipment sound recognition method mainly studies the transformer, cable, circuit breaker and other electrical main equipment, and lacks the sound recognition research of the sealed equipment in the power plant.The current commonly used classification recognition method adopts neural network recognition. Jiang et al. of Guangzhou Branch of Ultra-high Voltage Transmission Company proposed a sound recognition and power equipment fault warning method, system, terminal and medium, mainly using Kalman filter to extract the frequency band characteristics of sound data for sound recognition (Jiang Y, Wang R, Hong Z, Yan W, Xu H, Yuan H, Shi Y, Ye Z, Huang J, Li K, Lu G, Gao X, Li Z. Sound recognition and power equipment fault warning method, system, terminal and medium[P]. Guangdong Province: CN113314144A, 2021-08-27.); Li Zongzheng et al. of Shanghai Mai Nei Energy Technology Co., Ltd. proposed an intelligent monitoring system and monitoring method for power equipment based on sound recognition, which recognizes real-time audio according to tensorflow software, and alarms if the sound model identifies that there is no sound (Li Z, Lu Z, Liu F, Yan C, Wang Z, Xia Y, Zhang F, Ding T, Wang J, Dai J, Wang J, Wu Y, Wang X, Jiang H, Zhang X. An intelligent monitoring system and monitoring method for power equipment based on sound recognition[P]. Shanghai: CN110867959A, 2020-03-06.); Wu Qiang et al. of Sichuan Changhong Electric Appliance Co., Ltd. proposed a device abnormal sound recognition method and system based on convolutional neural network, which acquires effective interval characteristics according to the frequency domain difference characteristics of normal devices and abnormal devices, and uses convolutional neural network for recognition (Wu Q, Zhan H Y, Liu Y. Device abnormal sound recognition method and system based on convolutional neural network[P]. Sichuan Province: CN111076809B, 2021-08-31.); Xiao Xiangping of Jingzhou City Juye Transmission Machinery Co., Ltd. proposed a gear box automatic fault diagnosis method based on neural network and characteristic frequency band, which first establishes an amplitude-frequency diagram for the vibration acceleration of the device, and uses neural network to diagnose the fault probability of the secondary wheel (Xiao X, Chen L, Zhu F, Zhao J. Gear box automatic fault diagnosis method based on neural network and characteristic frequency band[P]. Hubei Province: CN110836770A, 2020-02-25.); Hu Sheng et al. of State Grid Corporation proposed a method and device for diagnosing generator discharge fault using noise characteristic frequency band, which arranges multiple noise measurement points on the generator, obtains the characteristic frequency band noise distribution curve according to the preset noise signal of each measurement point, and finds out whether the generator has discharge fault by judging whether there is an abnormal rising measurement point and whether the difference between the maximum value and the minimum value of the measurement point exceeds the threshold (Hu S, Hao J, Meng Z, Xu B, Zhen G, Wu X, Tang J, Sun B, Chen X, Peng J. A method and device for diagnosing generator discharge fault using noise characteristic frequency band[P]. Beijing: CN105866645B, 2018-10-19.); The above invention methods do not consider the air leakage detection of the device, and do not perform feature screening on all frequency bands. SUMMARY

[0004] The application provides an electrical equipment sound recognition method based on feature frequency band selection, which can recognize normal working condition, normal air leakage working condition and abnormal air leakage working condition of the electrical equipment simultaneously.

[0005] The application is characterized in that the Metropolis-Hastings sampling method is used to sample the sound frequency domain, the frequency band features are extracted through the logarithmic attenuation function and the partial least squares regression, and the neural network is built to recognize the sound, so that the electrical equipment sound monitoring method with higher efficiency and better effect is obtained by combination.

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

[0007] An electrical equipment sound recognition method based on feature frequency band selection comprises the following steps.

[0008] S1, sound time domain data acquisition and preprocessing, Fourier transform is used to convert to the frequency domain;

[0009] S2, the Metropolis-Hastings sampling method is used to randomly sample a set number of samples to establish a partial least squares regression model;

[0010] S3, the logarithmic attenuation function is used to filter the frequency band with the minimum absolute value weight of the regression coefficient;

[0011] S4, the frequency band with a set retention ratio is selected by using weighted sampling, and the corresponding frequency band with smaller cross-validation root mean square error is selected as the feature frequency band;

[0012] S5, a neural network is established to recognize the sound state.

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

[0014] The preprocessing is a length slicing and normalization processing on the acquired sound time domain data, and the specific steps are as follows.

[0015]

[0016] Wherein, l i is the sound signal of the i-th segment after the sound time domain data is divided, S c is the minimum time length of the sound information, L is the sound time domain data acquired by the sensor, and N is the number of slice segments.

[0017] The frequency domain signal is converted from the original sound signal to the frequency domain by using fast Fourier transform.

[0018]

[0019] wherein F k is the kth frequency band information of the sound.

[0020] Further, in step S2, the processed sound frequency domain signal is randomly sampled by using the Metropolis-Hastings sampling method and a partial least squares regression model is established, including the following steps:

[0021] S2.1, randomly sampling by using the Metropolis-Hastings sampling method;

[0022] S2.2, selecting a set number of samples to establish a partial least squares regression model:

[0023] The set number of samples are put into the modeling set, and the remaining samples are used as the prediction set, and a partial least squares regression model is established, and then the regression coefficients of the partial least squares regression model are solved.

[0024] Further, in step S2.1, for the sound frequency domain distribution G(θ), first assume a proposal distribution R(θ new |θ last ), randomly obtain the first sample θ1, then sample θ new from the proposal distribution, calculate the relative probability density η new of the first sample θ1and the sample θ new,1 , if η>1, θ new is taken as the next sample θ2; if η<1, a random number u is taken, when p(u)<η, θ new is taken as the next sample θ2; otherwise, continue to sample with θ1until the entire random sampling process is completed.

[0025] Further, the calculation of the relative probability density η new,a is as follows:

[0026]

[0027] wherein a is the sample number, θ a is the ath sample of the sound frequency domain distribution G(θ), P G (θ a |θ) is the probability of extracting θ a from the sound frequency domain distribution, P R (θ new |θ) is the probability of extracting θ new from the sound frequency domain distribution, ηnew,a the relative probability density of the first sample and the next sample θ new the relative probability density of the first sample and the next sample θ

[0028] Further, in step S3, the frequency band with the minimum absolute value weight of the regression coefficient is filtered by using a logarithmic attenuation function:

[0029] The logarithmic attenuation function is:

[0030] μ = -βlnq

[0031] wherein μ is the reserved frequency band ratio; β and q are both constants, which are default parameters artificially set.

[0032] Further, in the first sampling, all frequency bands participate in the calculation, and at this time, the reserved frequency band ratio is 1, and in the last sampling, the reserved ratio is μ.

[0033] Further, in step S4, the frequency band with the set reserved ratio is selected by using weighted sampling, and the corresponding frequency band with smaller cross-validation root mean square error is selected as the characteristic frequency band:

[0034] In each sampling, the frequency band variable with a ratio of μ is selected by using weighted sampling, partial least squares regression modeling is performed, and the cross-validation root mean square error of each sampling frequency band is calculated:

[0035]

[0036] wherein RM is the cross-validation root mean square error, y is the true label value, y z is the predicted label value obtained by sampling the frequency band, and A is the sample number.

[0037] Further, in step S5, the neural network is a three-layer neural network, and the details are as follows:

[0038] The input layer is the selected characteristic frequency band; the output of the two hidden layers is the electrical equipment state, the error is obtained by using forward propagation, and the weight is updated in reverse through the partial derivative and the learning rate;

[0039] The neural network layer propagation is:

[0040] s = ωx + b

[0041] wherein s is the number of neurons of the next layer network, x is the number of neurons of the previous layer network, b is the bias term, and ω is the weight.

[0042] Further, the weight update formula is:

[0043]

[0044] Wherein, omega is the updated weight, loss is the neural network error, and alpha is the learning rate.

[0045] Compared with the prior art, the present application has the advantages of:

[0046] The present application provides an electrical equipment sound recognition method based on feature frequency band selection, effectively solving the problems of insufficient practicability and poor reliability of the current sound monitoring technology through gas monitoring and physical detection, using the sampling and regression method to extract the frequency band characteristics of the equipment, and achieving accurate recognition effect based on frequency band training. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The present application provides an electrical equipment sound recognition method based on feature frequency band selection, effectively solving the problems of insufficient practicability and poor reliability of the current sound monitoring technology through gas monitoring and physical detection, using the sampling and regression method to extract the frequency band characteristics of the equipment, and achieving accurate recognition effect based on frequency band training.

[0048] Figure 2 The present application provides an electrical equipment sound recognition method based on feature frequency band selection, effectively solving the problems of insufficient practicability and poor reliability of the current sound monitoring technology through gas monitoring and physical detection, using the sampling and regression method to extract the frequency band characteristics of the equipment, and achieving accurate recognition effect based on frequency band training.

[0049] Figure 3 The present application provides an electrical equipment sound recognition method based on feature frequency band selection, effectively solving the problems of insufficient practicability and poor reliability of the current sound monitoring technology through gas monitoring and physical detection, using the sampling and regression method to extract the frequency band characteristics of the equipment, and achieving accurate recognition effect based on frequency band training. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] Embodiment 1:

[0052] An electrical equipment sound recognition method based on feature frequency band selection, as shown in the figure, includes the following steps: Figure 1

[0053] S1, sound time domain data acquisition and preprocessing, converting to frequency domain by Fourier transform;

[0054] The sound time domain data is a sound time domain signal obtained from a sound sensor installed on the electrical equipment;

[0055] The preprocessing is to slice and normalize the acquired sound time domain data according to the time length, specifically as follows:

[0056]

[0057] Wherein, l i ​S is a sound signal of an i-th segment of sound time domain data c is a minimum time length of reserving sound information, L is sound time domain data acquired by a sensor, and N is a number of segments.

[0058] The frequency domain signal is converted from an original sound signal to a frequency domain by using a fast Fourier transform.

[0059]

[0060] wherein F k is a k-th frequency band information of sound.

[0061] S2, a Metropolis-Hastings sampling method is used to randomly sample a set number of samples to establish a partial least squares regression model, as shown in the following formula: Figure 2 The method comprises the following steps:

[0062] S2.1, a Metropolis-Hastings sampling method is used to randomly sample:

[0063] For a sound frequency domain distribution G(θ), a proposal distribution R(θ new |θ last ) is first assumed, a first sample θ1 is randomly obtained, and θ new is sampled from the proposal distribution, the relative probability density η new of the first sample θ1 and the sample θ new,1 is calculated, if η>1, θ new is taken as a next sample θ2; if η<1, a random number u is taken, when p(u)<η, θ new is taken as a next sample θ2; otherwise, sampling is continued with θ1 until the whole random sampling process is completed.

[0064] The calculation of the relative probability density η new,a is as follows:

[0065]

[0066] wherein a is a sample serial number, θ a is an a-th sample of the sound frequency domain distribution G(θ), P G (θ a |θ) is a probability of sampling θ a from the sound frequency domain distribution, P R (θ new |θ) is a probability of sampling θ new from the sound frequency domain distribution, and η new,a is a relative probability density of the a-th sample and a next sample θ new . A partial least squares regression model;

[0067] S2.2, select a set number of samples to establish a partial least squares regression model:

[0068] Put the set number of samples into the modeling set, and the remaining samples as the prediction set, establish a partial least squares regression model, and then solve the partial least squares regression model.

[0069] S3, filter the frequency band with the smallest absolute value weight of the regression coefficient by using a logarithmic decay function;

[0070] The logarithmic decay function is:

[0071] μ = -βlnq

[0072] Wherein, μ is the reserved frequency band ratio, β and q are both constants, and are artificial set values;

[0073] In the first sampling, all frequency bands participate in the calculation, at this time the reserved frequency band ratio is 1, and the reserved ratio in the last sampling is μ.

[0074] S4, select the frequency band with a set reserved ratio by using weighted sampling, and select the corresponding frequency band with smaller cross-validation root mean square error as the characteristic frequency band;

[0075] In each sampling, the frequency band variable with a ratio of μ is selected by using weighted sampling, partial least squares regression modeling is performed, and the cross-validation root mean square error of each sampling frequency band is calculated:

[0076]

[0077] Wherein, RM is the cross-validation root mean square error, y is the true label value, y z is the predicted label value obtained by sampling the frequency band, and A is the number of samples.

[0078] S5, establish a neural network to identify the sound state;

[0079] As Figure 3 shown, the neural network is a three-layer neural network, specifically as follows:

[0080] The input layer is the selected characteristic frequency band; two layers of hidden layers, and the output is the electrical equipment state. The error is calculated by forward propagation, and the weight is updated in reverse by partial derivative and learning rate;

[0081] The neural network layer propagation is:

[0082] s = ωx + b

[0083] Wherein, s is the number of neurons of the next layer network, x is the number of neurons of the previous layer network, b is the bias term, and ω is the weight;

[0084] The weight update formula is:

[0085]

[0086] where ω' is the updated weight, loss is the neural network error, and a is the learning rate.

[0087] In this embodiment, 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 and frequency domain conversion processing, sound frequency domain samples are obtained, and the data dimension is [960, 96000]. 1831 characteristic frequency bands are extracted by using the Metropolis-Hastings sampling method and logarithmic attenuation.

[0088] The input layer of the neural network is set to 1831 nodes, the hidden layer is set to 20 neurons, the predicted label value is output by using the full connection layer, and after 100 times of training, the overall accuracy is 99.79%, and the test set results are shown in Table 1.

[0089] Table 1

[0090] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 99.61％ 99.82％ 99.82％

[0091] Embodiment 2:

[0092] 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 and frequency domain conversion processing, sound frequency domain samples are obtained, and the data dimension is [960, 96000]. 1179 characteristic frequency bands are extracted by using the Metropolis-Hastings sampling method and logarithmic attenuation.

[0093] The input layer of the neural network is set to 1179 nodes, the hidden layer is set to 20 neurons, the predicted label value is output by using the full connection layer, and after 100 times of training, the overall accuracy is 99.06%, and the test set results are shown in Table 2.

[0094] Table 2

[0095] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 99.1％ 99.4％ 98.8％

[0096] Embodiment 3:

[0097] 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 and frequency domain conversion processing, sound frequency domain samples are obtained, and the data dimension is [960, 96000]. 2843 characteristic frequency bands are extracted by using the Metropolis-Hastings sampling method and logarithmic attenuation.

[0098] The neural network input layer is set to 2843 nodes, the hidden layer is set to 10 neurons, the predicted label value is output by using a full connection layer, after 100 times of training, the overall accuracy is 99.79%, and the test set result is shown in Table 3.

[0099] Table 3

[0100] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the invention 99.7％ 99.3％ 99.6％

[0101] The above identification and tracking method is combined as a preferred embodiment of the present application, but the embodiments of the present application are not limited by the above examples, and any modification, modification, substitution, combination, simplification made without departing from the spirit and principles of the present application should be equivalent replacement mode, and should be included in the protection scope of the present application.

Claims

1. A method for sound recognition of electrical equipment based on characteristic frequency band selection, characterized in that, Includes the following steps: S1. Acquisition and preprocessing of sound time-domain data, followed by conversion to the frequency domain using Fourier transform; S2. Using the Metropolis-Hastings sampling method, a set number of samples are randomly sampled to establish a partial least squares regression model; S3. Use the logarithmic decay function to filter the frequency band with the smallest absolute value weight of the regression coefficient; S4. Select the frequency band with the set retention ratio using weighted sampling, and select the corresponding frequency band as the characteristic frequency band based on the root mean square error of the cross-validation. In each sampling, a frequency band variable with a weighted sampling selection ratio of μ is used for partial least squares regression modeling, and the root mean square error of cross-validation for each sampling frequency band is calculated: in, RM To interactively verify the root mean square error, y The actual label value. y z The predicted label value is obtained from the sampled frequency band. A The number of samples; S5. Establish a neural network to identify the state of sound.

2. The method for electrical equipment sound recognition based on characteristic frequency band selection 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 time-domain data by duration and normalizing it, as detailed below: in, i The first segment of the sound time-domain data i fragmented sound signals, S c It is the minimum time length for retaining sound information. L For the sound time-domain data acquired by the sensor, N The number of segments that can be cut; Frequency domain signals are generated by converting the original sound signal to the frequency domain using Fast Fourier Transform. in, F k For the sound k Frequency band information.

3. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 1, characterized in that, In step S2, the processed audio frequency domain signal is randomly sampled using the Metropolis-Hastings sampling method, and a partial least squares regression model is established, including the following steps: S2.

1. Random sampling is performed using the Metropolis-Hastings sampling method; S2.

2. Select a set number of samples to establish a partial least squares regression model: A set number of samples are entered into the modeling set, and the remaining samples are used as the prediction set. A partial least squares regression model is established, and then the regression coefficients are solved for the partial least squares regression model.

4. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 3, characterized in that, In step S2.1, regarding the audio frequency domain distribution... G ( θ First, assume a proposed distribution. R ( θ new | θ last ), randomly obtain the first sample θ 1. Then sample from the proposed distribution. θ new Calculate the first sample θ 1 and sample θ new relative probability density η new,1 ,if η >1, θ new Then it will be used as the next sample. θ 2; if η <1, generate a random number u ,when p ( u )< η hour, θ new Then it will be used as the next sample. θ 2; otherwise, continue with θ 1. Sampling is carried out until the entire random sampling process is completed.

5. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 4, characterized in that, relative probability density The specific calculation is as follows: in, a The sample number. For the distribution of sound frequency domain G ( θ ) No. a One sample, To extract from the audio frequency domain distribution The probability, To extract from the audio frequency domain distribution The probability, η new,a For the first a The first sample and the next sample θ new The relative probability density.

6. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 1, characterized in that, In step S3, the frequency band with the smallest absolute value weight of the regression coefficient is filtered out using the logarithmic decay function: The logarithmic decay function is: in, μ The proportion of frequency bands to be reserved; β and q All of these are constants, default parameters set manually.

7. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 6, characterized in that, In the first sampling, all frequency bands are involved in the calculation, and the proportion of frequency bands retained at this time is 1. In the last sampling, the proportion retained is... μ .

8. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 1, characterized in that, In step S5, the neural network is a three-layer neural network, as detailed below: The input layer consists of the selected feature frequency bands; the output of the two hidden layers is the electrical equipment status. The error is calculated using forward propagation, and the weights are updated in reverse using partial derivatives and the learning rate. Inter-layer propagation in a neural network is as follows: in, s This represents the number of neurons in the next layer of the network. x This represents the number of neurons in the previous layer of the network. b For deviation terms, ω For weights.

9. The method for electrical equipment sound recognition based on characteristic frequency band selection according to claim 8, characterized in that, The weight update formula is: in, ω` For the updated weights, loss For neural network error, α This is the learning rate.

Images