Circuit breaker mechanical fault diagnosis method based on sound signals

Abstract

The invention provides a circuit breaker mechanical fault diagnosis method based on a sound signal, and relates to the technical field of electric power operation and maintenance and equipment fault detection, and the method comprises the steps: collecting a sound signal generated by a to-be-diagnosed circuit breaker in the switching-on and switching-off process; preprocessing the sound signal to obtain a preprocessed sound signal containing effective information; performing denoising processing on the preprocessed sound signal based on an adversarial auto-encoder to obtain a denoised sound signal; performing time-frequency feature extraction on the de-noised sound signal based on short-time Fourier transform to obtain a time-frequency spectrogram; performing zero-crossing rate and kurtosis detection on the de-noised sound signal to obtain a zero-crossing rate feature and a kurtosis feature; and according to the time-frequency spectrogram, the zero-crossing rate feature and the kurtosis feature, diagnosing the mechanical fault of the to-be-diagnosed circuit breaker. According to the scheme, the accuracy of a fault diagnosis result can be improved while the fault diagnosis cost of the circuit breaker is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power operation and maintenance and equipment fault detection, and in particular to a circuit breaker mechanical fault diagnosis method based on acoustic signals. Background Art

[0002] Circuit breakers play a key protective role in power systems. Circuit breakers are used to protect power system equipment from damage caused by abnormal conditions such as overload and short circuit, ensuring system safety and stable power supply. However, due to the high frequency of operation and current shocks that circuit breakers experience during long-term use, their internal mechanical parts are prone to wear, aging, and damage. If a circuit breaker fails, it may cause serious equipment damage or even lead to large-scale power failures. Therefore, it is of great significance to detect and maintain circuit breakers.

[0003] Traditional methods for fault detection of circuit breakers are usually based on temperature, vibration, electrical and other measurement data, usually detecting the current, voltage or heat of the circuit breaker to determine its health status. However, although these methods can achieve fault detection of circuit breakers to a certain extent, they usually require the installation of expensive sensors or special equipment to detect various physical parameters of the circuit breaker, which is costly. Moreover, the gradual accumulation of internal faults in circuit breakers in traditional methods is often ignored and can only be detected when significant changes occur. In addition, traditional methods have higher restrictive requirements for physical conditions, and parameters such as current and temperature are easily affected by external interference, especially in harsh environments or complex power networks, and the fault diagnosis results are low. Summary of the invention

[0004] In view of this, in order to address the above shortcomings, it is necessary to propose a circuit breaker mechanical fault diagnosis method based on acoustic signals to reduce the circuit breaker fault diagnosis cost while improving the accuracy of fault diagnosis results.

[0005] The present invention provides a circuit breaker mechanical fault diagnosis method based on acoustic signals, comprising:

[0006] Collect the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process;

[0007] Preprocessing the sound signal to obtain a preprocessed sound signal containing effective information;

[0008] Performing denoising on the preprocessed sound signal based on an adversarial autoencoder to obtain a denoised sound signal;

[0009] Extracting time-frequency features of the de-noised sound signal based on short-time Fourier transform to obtain a time-frequency spectrum;

[0010] Performing zero-crossing rate and kurtosis detection on the denoised sound signal respectively to obtain a zero-crossing rate feature and a kurtosis feature;

[0011] The mechanical fault of the circuit breaker to be diagnosed is diagnosed according to the time-frequency spectrum diagram, the zero-crossing rate feature and the kurtosis feature.

[0012] Preferably, the preprocessing of the sound signal includes:

[0013] Performing frame processing on the sound signal to obtain a frame signal; wherein there is overlap between adjacent frames during the frame processing;

[0014] Performing windowing processing on each frame of the framed signal to obtain a frame matrix;

[0015] Effective feature extraction is performed on the frame matrix to obtain the preprocessed sound signal.

[0016] Preferably, the performing windowing processing on each frame of the framed signal includes:

[0017] For each frame signal x k [n], perform windowing processing by the following calculation formula to obtain N windowed signals;

[0018]

[0019] in, It is used to represent the nth windowed signal obtained by the kth frame signal, x k [n] is used to represent the kth frame signal of the framed signal, w[n] is used to represent the window function, N is the frame length, 0≤n<N;

[0020] Convert each frame signal into a row vector and construct the following frame matrix:

[0021]

[0022] Wherein, X is used to represent the frame matrix, and K is used to represent the number of frames.

[0023] Preferably, the effective feature extraction for the frame matrix includes:

[0024] For each frame signal, the total energy value of the corresponding frame signal is calculated by the following calculation formula:

[0025]

[0026] Among them, E k Used to characterize the total energy value corresponding to the k-th frame signal, Used to represent the nth windowed signal corresponding to the kth frame signal;

[0027] Each frame signal whose total energy value is greater than a preset threshold is determined as the preprocessed sound signal.

[0028] Preferably, the process of constructing the adversarial autoencoder includes:

[0029] Preprocessing the historical preprocessed sound signal data to obtain preprocessed sample data; wherein the preprocessed sample data conforms to the input format of the model;

[0030] Mapping the preprocessed sample data to a latent space based on an encoder to generate a low-dimensional latent representation, and remapping the latent representation to an original data space based on a decoder to obtain first sample training data without noise;

[0031] Performing autoencoder training; specifically, the steps include: mapping the noise preprocessed sample data to a latent space to obtain a latent representation, and converting the latent representation to the original data space based on a decoder to obtain a second sample training data containing noise; constructing a reconstruction loss based on the difference between the second sample training data containing noise and the first sample training data containing no noise, and performing training by minimizing the reconstruction loss;

[0032] Conduct adversarial training; specifically, the method includes: selecting a suitable prior distribution as the target distribution of the latent space; inputting the preprocessed sample data as real data into the discriminator, and generating a potential representation through the encoder using the preprocessed sample data with added noise, and inputting the potential representation into the discriminator as forged data; optimizing the discriminator through the cross entropy loss, so that the potential representation generated by the encoder is as close to the prior distribution as possible;

[0033] The autoencoder training and the adversarial training are alternately iterated to obtain an adversarial autoencoder model for denoising the preprocessed sound signal.

[0034] Preferably, the extracting time-frequency features of the denoised sound signal based on short-time Fourier transform comprises:

[0035] At each time point t on the time axis, a window function is used to intercept the local signal;

[0036] Perform Fourier transform on each intercepted local signal to obtain the frequency characteristics of each local signal;

[0037] The time-frequency spectrum is constructed by using the frequency characteristics of the local signal intercepted at each time point on the time axis.

[0038] Preferably, the extracting time-frequency features of the denoised sound signal based on short-time Fourier transform comprises:

[0039] For a given signal x(t), the following calculation formula is used to extract time-frequency features:

[0040]

[0041] Among them, X(t,f) is used to characterize the frequency characteristics corresponding to the time point t, w(τ-t) is used to characterize the window function, and f is the frequency.

[0042] Preferably, performing kurtosis detection on the denoised sound signal comprises:

[0043] For each frame signal in the preprocessed sound signal, the kurtosis is calculated using the following calculation formula:

[0044]

[0045] Among them, Kurtosis is used to characterize the kurtosis characteristics, x i is the i-th signal in the current frame signal of the preprocessed sound signal after denoising, n is the number of data points, μ is the mean of the data, and σ is the variance.

[0046] Preferably, performing zero-crossing rate detection on the denoised sound signal comprises:

[0047] For each frame signal in the preprocessed sound signal, the zero-crossing rate is calculated using the following calculation formula:

[0048]

[0049] Among them, ZCR is used to represent the zero-crossing rate, N is used to represent the frame length, and x[n] is used to represent the nth signal of the current frame signal in the denoised preprocessed signal.

[0050] Preferably, the diagnosing the mechanical fault of the circuit breaker to be diagnosed according to the time-frequency spectrum diagram, the zero-crossing rate feature and the kurtosis feature includes:

[0051] Inputting the one-dimensional zero-crossing rate feature and the kurtosis feature into a one-dimensional convolutional neural network of a pre-trained classification model to obtain a one-dimensional first feature vector;

[0052] Inputting the two-dimensional time-frequency spectrum graph into a two-dimensional convolutional neural network of a pre-trained classification model to obtain a two-dimensional second feature vector;

[0053] The first eigenvector and the second eigenvector are concatenated and input into a fully connected layer to obtain a mechanical fault diagnosis result of the circuit breaker.

[0054] It can be seen from the above technical scheme that when diagnosing the mechanical fault of the circuit breaker based on this scheme, the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process is first collected, and then the sound signal is processed to obtain a pre-processed sound signal containing effective information; further, the pre-processed sound signal is denoised based on the adversarial autoencoder to obtain a denoised sound signal; then, the denoised sound signal is extracted from the time-frequency feature based on the short-time Fourier change, and the zero-crossing rate and kurtosis of the denoised sound signal are detected at the same time, and finally, the fault diagnosis of the circuit breaker to be diagnosed can be realized based on the obtained time-frequency spectrum, zero-crossing rate feature and kurtosis feature. It can be seen that the detection equipment in this scheme only needs to use a voiceprint sensor, and the voiceprint detection sensor has a low cost and is easy to install and maintain, which can greatly reduce the detection and operation and maintenance costs. In addition, this scheme denoises the collected signal based on the adversarial encoder, which can eliminate the influence of background noise, thereby improving the effect of subsequent feature extraction, and then improving the accuracy of the model and the accuracy of fault diagnosis. At the same time, the extraction of time-frequency features in this scheme can find out all the closing feature positions. Based on the characteristics that kurtosis and zero-crossing rate are very sensitive to abnormal values, the detection of kurtosis and zero-crossing rate can more effectively and accurately detect the characteristic changes of circuit breaker faults, making the fault diagnosis results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of a circuit breaker mechanical fault diagnosis method based on acoustic signals provided by an embodiment of the present invention.

[0056] Figure 2 This is a short-time Fourier change diagram under normal conditions.

[0057] Figure 3 It is a short-time Fourier change diagram under the fault state of too low opening and closing speed.

[0058] Figure 4 It is the changing trend of kurtosis under normal conditions.

[0059] Figure 5 It is a kurtosis change trend under the fault state of too low opening and closing speed.

[0060] Figure 6 It is the changing trend of the zero-crossing rate under normal conditions.

[0061] Figure 7 It is a trend of zero-crossing rate change under the fault state of too low opening and closing speed.

[0062] Figure 8 A schematic diagram of fault diagnosis based on time-frequency spectrum, zero-crossing rate characteristics and kurtosis characteristics. DETAILED DESCRIPTION

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0064] Sound diagnosis is a new type of non-invasive detection method. When the circuit breaker is in normal or faulty operation, it will produce specific sound wave characteristics. These sound wave characteristics reflect the status of the mechanical parts of the circuit breaker, which is called the soundprint of the circuit breaker. The soundprint detection method is based on the sound generated during the mechanical movement inside the circuit breaker. By analyzing these acoustic signals, the early characteristics of the circuit breaker failure can be found. Compared with traditional methods, soundprint detection has the following advantages: non-contact detection, soundprint detection does not need to be directly connected to the electrical circuit or mechanical structure of the circuit breaker, and can collect sound signals over a long distance, avoiding complex wiring and safety hazards. Low cost, the acoustic sensor required for soundprint detection is usually low cost, easy to install and maintain, and suitable for promotion in large-scale power systems. Real-time and high sensitivity, sound waves are extremely sensitive to changes in the status of mechanical parts. Soundprint detection can detect the opening and closing process of the circuit breaker in real time and capture early mechanical failure characteristics. Strong adaptability, soundprint detection performs well in complex environments, can adapt to circuit breakers of different models and sizes, and is not easily affected by external electrical interference.

[0065] Based on this, Figure 1 As shown, this scheme considers proposing a circuit breaker mechanical fault diagnosis method based on acoustic signals, which may include the following steps:

[0066] Step 10: Collect the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process;

[0067] Step 20: preprocessing the sound signal to obtain a preprocessed sound signal containing effective information;

[0068] Step 30: De-noising the pre-processed sound signal based on the adversarial autoencoder to obtain a de-noised sound signal;

[0069] Step 40: extracting time-frequency features of the denoised sound signal based on short-time Fourier transform to obtain a time-frequency spectrum;

[0070] Step 50: performing zero-crossing rate and kurtosis detection on the denoised sound signal to obtain a zero-crossing rate feature and a kurtosis feature;

[0071] Step 60: diagnose the mechanical fault of the circuit breaker to be diagnosed according to the time-frequency spectrum diagram, the zero-crossing rate feature and the kurtosis feature.

[0072] In this embodiment, when performing fault diagnosis of the circuit breaker, the only detection equipment required is the voiceprint sensor, and the voiceprint detection sensor has a low cost, is easy to install and maintain, and can greatly reduce the detection and operation and maintenance costs. Moreover, when the circuit breaker is opened and closed, a sound signal is generated in this process, and the sound signal contains fault information, but there is external background noise in the collected signal. Therefore, this scheme considers denoising the collected signal based on the adversarial encoder, which can eliminate the influence of background noise, thereby improving the effect of subsequent feature extraction, and then improving the accuracy of the model and the accuracy of fault diagnosis. In addition, the short-time Fourier transform can effectively extract the time-frequency characteristics of the fault signal, which is suitable for processing non-stationary signals, and can extract the sound characteristics of the closing moment. By observing the short-time Fourier transform diagram, all the closing feature positions can be seen. Based on the characteristics that the kurtosis and zero-crossing rate are very sensitive to abnormal values, the characteristic changes of the circuit breaker fault can be detected more effectively and accurately through the detection of the kurtosis and zero-crossing rate, so that the result of fault diagnosis is more accurate.

[0073] For step 10, the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process is collected;

[0074] In this step, consider collecting the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process, for example, by combining a magnetic base and a microphone to collect the sound signal, and save the collected sound signal. When collecting data, the collection strategy is usually to collect 30 seconds of data every minute, and perform multiple opening and closing operations during the collection process.

[0075] For step 20, the sound signal is preprocessed to obtain a preprocessed sound signal containing effective information;

[0076] In this step, the collected sound signal is processed by framing, and the useful signal part is extracted for storage, and the data segments without valid information are discarded. Specifically, step 20 can be implemented as follows:

[0077] Step S21: performing frame processing on the sound signal to obtain a frame signal; wherein there is overlap between adjacent frames during the frame processing;

[0078] Step S22: performing windowing processing on each frame of the framed signal to obtain a frame matrix;

[0079] Step S23: extract effective features from the frame matrix to obtain a preprocessed sound signal.

[0080] When processing the sound signal in frames in step S21, the frame length of each frame can be set to 3*sr, where sr is the sampling rate of the sensor. The frame shift is used to characterize the distance between the starting positions of each frame, which is usually set to about 50% of the frame length to ensure that adjacent frames have a certain overlap, thereby retaining the continuous characteristics of the signal. Therefore, the frame shift is set to 1.5*sr. Assuming the total length of the signal is L, the frame length is N, the frame shift is M, and the number of frames is K, the calculation formula for the number of frames K is as follows:

[0081]

[0082] This formula ensures that the last frame of the signal does not exceed the total length L while covering the entire signal.

[0083] In order to reduce the signal discontinuity at the frame boundary, consider performing windowing processing on the signal in step S22. k [n], the windowing process can be performed by the following calculation formula, and the k-th frame signal after windowing can be expressed as:

[0084]

[0085] in, It is used to represent the nth windowed signal obtained by the kth frame signal, x k [n] is used to represent the kth frame signal of the framed signal, w[n] is used to represent the window function, N is the frame length, 0≤n<N;

[0086] For the window function w[n], common window functions are: rectangular window: w[n] = 1; Hamming window: Hanning Window: Any window function can be used to implement this solution.

[0087] After obtaining the windowing results of each frame, each frame signal is converted into a row vector to construct the following frame matrix:

[0088]

[0089] Among them, X is used to represent the frame matrix, and K is used to represent the number of frames. That is, each row in the matrix represents the result of windowing a frame of signal.

[0090] Furthermore, since a loud noise will be generated during the process of opening and closing the gate, that is, normally, the opening and closing action will only be performed once in a long period of time, that is, a lot of the collected data is invalid data, that is, the corresponding signal that the opening and closing operation is not performed, and it is obviously unreasonable to perform fault diagnosis on invalid signals. Therefore, consider performing effective feature extraction on the frame matrix through step S23 to reduce the interference of noise in fault diagnosis, while reducing the data dimension and improving the execution efficiency of the processor. Specifically, step S23 can be implemented in the following way:

[0091] For each frame signal, the total energy value of the corresponding frame signal is calculated by the following calculation formula:

[0092]

[0093] Among them, E k Used to characterize the total energy value corresponding to the k-th frame signal, Used to represent the nth windowed signal corresponding to the kth frame signal;

[0094] Each frame signal whose total energy value is greater than a preset threshold is determined as a preprocessed sound signal.

[0095] In this embodiment, high-intensity sound waves will occur during the opening and closing process. The total energy value of each frame signal is calculated, and it is determined whether the total energy value of each frame signal is greater than 50; if it is greater than 50, it means that the frame signal corresponds to the sound signal of the opening and closing operation, if not, it means that it is not. In this way, by retaining the frame signals with a total energy signal value greater than 50 and discarding the frame signals with a total energy signal value less than or equal to 50, the purpose of screening and extracting effective features in the frame matrix can be achieved.

[0096] For step 30, the preprocessed sound signal is denoised based on the adversarial autoencoder to obtain a denoised sound signal;

[0097] In this step, the autoencoder is an unsupervised learning model, which usually consists of two parts: encoder and decoder. The encoder is responsible for mapping the high-dimensional raw data to the low-dimensional latent space, while the decoder converts the latent representation back to the representation of the original data space to form a reconstruction result. The training goal of the autoencoder is to minimize the reconstruction error so that the reconstructed data is as close to the input data as possible, so that the model can learn how to represent and compress the input data.

[0098] However, ordinary autoencoders cannot control the distribution of the latent space, which means that the latent representation generated by the encoder may not conform to a specific prior distribution (such as a Gaussian distribution). Such defects will cause the autoencoder to perform poorly when generating new data or processing noise-free data. Therefore, an adversarial training mechanism is introduced to make the latent space more structured.

[0099] Specifically, when building an adversarial autoencoder, it can be achieved in the following ways:

[0100] Preprocessing the historical preprocessed sound signal data to obtain preprocessed sample data; wherein the preprocessed sample data conforms to the input format of the model;

[0101] Mapping the preprocessed sample data to a latent space based on the encoder to generate a low-dimensional latent representation, and remapping the latent representation to the original data space based on the decoder to obtain first sample training data without noise;

[0102] Performing autoencoder training; specifically, the steps include: mapping the noise preprocessed sample data to a latent space to obtain a latent representation, and converting the latent representation to the original data space based on a decoder to obtain a second sample training data containing noise; constructing a reconstruction loss based on the difference between the second sample training data containing noise and the first sample training data containing no noise, and performing training by minimizing the reconstruction loss;

[0103] Conduct adversarial training; specifically, the method includes: selecting a suitable prior distribution as the target distribution of the latent space; inputting the preprocessed sample data as real data into the discriminator, and generating a potential representation through the encoder using the preprocessed sample data with added noise, and inputting the potential representation into the discriminator as forged data; optimizing the discriminator through the cross entropy loss, so that the potential representation generated by the encoder is as close to the prior distribution as possible;

[0104] The autoencoder training and the adversarial training are alternately iterated to obtain an adversarial autoencoder model for denoising the preprocessed sound signal.

[0105] In this embodiment, when constructing the adversarial autoencoder, the historical preprocessed sound signal data is first subjected to basic preprocessing, such as normalization. The data is then prepared into a format suitable for model input. The data is usually speech or other types of high-dimensional data, which can be compressed and represented by the autoencoder.

[0106] Then the autoencoder is constructed; the encoder maps the input data to the latent space, that is, extracts the key features of the input and generates a low-dimensional latent representation. The output of the encoder is a latent representation vector, which has no specific distribution requirements in traditional autoencoders. The decoder remaps the latent representation to the original data space to reconstruct the input data and generate an output that is as similar to the input as possible.

[0107] Further training of the autoencoder is performed; the noisy data is input into the autoencoder, and the encoder maps the input data to the latent space to obtain a latent representation. The decoder converts the latent representation into the original data space and generates the reconstructed output data. The reconstruction loss is calculated by calculating the difference between the reconstructed data and the original noise-free data, such as the mean square error MSE or L1 loss. This loss is the optimization goal to ensure that the autoencoder can effectively restore the input. Minimize the reconstruction loss and adjust the parameters of the encoder and decoder to make the reconstructed data as close to the original data as possible.

[0108] Further adversarial training is performed; first, a suitable prior distribution is selected as the target distribution of the latent space. The collected samples are taken as real samples and these samples are input into the discriminator as real data. After adding noise to the input data, the latent representation is generated through the encoder, and these representations are input into the discriminator as fake data. The task of the discriminator is to distinguish between real latent samples and fake latent representations. The discriminator is optimized through cross entropy loss or other adversarial losses to improve its classification accuracy. The adversarial loss of the encoder is optimized so that the latent representation generated by the encoder is as close to the prior distribution as possible, thereby confusing the discriminator. In this way, the encoder can generate more structured latent representations.

[0109] Finally, continuous iterative training is performed, with the reconstruction phase and the adversarial phase alternating. First, the encoder is reconstructed to minimize the reconstruction loss; then the adversarial training phase is entered, so that the discriminator and encoder optimize their respective goals. Through alternating optimization, the potential representation generated by the encoder gradually approaches the prior distribution, and at the same time, the data can be effectively reconstructed, and the adversarial autoencoder model can be used to achieve data denoising.

[0110] For step 40, time-frequency features of the de-noised sound signal are extracted based on short-time Fourier transform to obtain a time-frequency spectrum;

[0111] In this step, consider extracting the time-frequency features of the denoised audio signal based on short-time Fourier transform, which can be specifically achieved in the following way:

[0112] At each time point t on the time axis, a window function is used to intercept the local signal;

[0113] Perform Fourier transform on each intercepted local signal to obtain the frequency characteristics of each local signal;

[0114] The time-frequency spectrum is constructed by using the frequency characteristics of the local signal intercepted at each time point on the time axis.

[0115] For example, step 40 may extract time-frequency features for a given signal x(t) using the following calculation formula:

[0116]

[0117] Among them, X(t,f) is used to characterize the frequency characteristics corresponding to the time point t, w(τ-t) is used to characterize the window function, and f is the frequency.

[0118] Short-time Fourier transform is a signal processing technology used to analyze the time-frequency characteristics of non-stationary signals. This solution is based on short-time Fourier transform. By dividing the signal into multiple short-time windows on the time axis, Fourier transform is performed on each window to obtain the spectrum information of the signal in each time period. That is, the window function slides on the time axis, and the time-frequency spectrum can be generated by Fourier transforming the signal in each window.

[0119] For step 50, the denoised sound signal is subjected to zero-crossing rate and kurtosis detection respectively to obtain a zero-crossing rate feature and a kurtosis feature;

[0120] Kurtosis is one of the statistical characteristics that describes the characteristics of data distribution. It is used to measure the peak degree and tail thickness of data distribution. Kurtosis is particularly suitable for identifying non-Gaussian features, judging the outliers of data, and analyzing the stability of signals. Kurtosis is the fourth-order moment of data distribution, which is used to measure the steepness of the peak and tail of the distribution curve near the mean. For example, given a set of data X = {x1, x2, ..., x n}, the formula for kurtosis is In the kurtosis formula, (x i -μ) 4 The data points that deviate far from the mean have a greater impact on the kurtosis value, so the kurtosis can reflect the tail weight and peak degree of the data distribution. In the case of high kurtosis, there are more discrete points, and it can be judged whether the circuit breaker is in normal working condition. If there is an abnormality, it will be reflected in the change of kurtosis.

[0121] Zero-crossing rate is an important feature in signal processing. It indicates the number of times the waveform of a signal passes through zero value within a certain time window, that is, the frequency at which the signal changes from positive value to negative value, or from negative value to positive value. Zero-crossing rate is mainly used to analyze the frequency characteristics and noise characteristics of a signal. For example, for a discrete signal x[n], the zero-crossing rate is defined as the number of times the signal changes its sign within a fixed time window. Given a window of length N, the zero-crossing rate can be calculated by Calculated.

[0122] The zero-crossing rate can be used to determine the voice endpoint. The zero-crossing rate of voice signals and silent segments usually has a significant difference. The zero-crossing rate of silent segments is low, while the zero-crossing rate of voice segments is high. Through zero-crossing rate detection, voice and non-voice signals can be accurately distinguished. Specifically, the zero-crossing rate can be used to determine whether the signal is in the process of opening and closing the switch. When the zero-crossing rate increases rapidly, it means that the circuit breaker is in the process of opening and closing the switch, and the zero-crossing rate feature at this time can be proposed. When a circuit breaker fails, for example, when the opening and closing speed is too low, the change in the zero-crossing rate will decrease. By extracting the zero-crossing rate feature, the fault can be diagnosed.

[0123] In step 60 , the mechanical fault of the circuit breaker to be diagnosed is diagnosed according to the time-frequency spectrum, the zero-crossing rate feature and the kurtosis feature.

[0124] like Figure 8 As shown, in this step, when performing fault diagnosis, firstly, the one-dimensional zero-crossing rate feature and the kurtosis feature are input into the one-dimensional convolutional neural network of the pre-trained classification model to obtain a one-dimensional first feature vector; then the two-dimensional time-frequency spectrum is input into the two-dimensional convolutional neural network of the pre-trained classification model to obtain a two-dimensional second feature vector; finally, the first feature vector and the second feature vector are concatenated and input into the fully connected layer to obtain the mechanical fault diagnosis result of the circuit breaker.

[0125] like Figure 2 and Figure 3 The short-time Fourier change diagram in the normal state and the short-time Fourier change diagram in the opening and closing speed too low fault state are shown respectively. Figure 4 and Figure 5 The following are the changing trends of the kurtosis under normal conditions and the changing trends of the kurtosis under the fault state of too low opening and closing speed. Figure 6 and Figure 7 They are the changing trend of zero-crossing rate under normal conditions and the changing trend of zero-crossing rate under the fault state of too low opening and closing speed. It can be seen from the comparison diagrams that there are obvious differences in the short-time Fourier change diagram, kurtosis change trend and zero-crossing rate change trend under normal conditions and fault conditions. Therefore, the fault diagnosis of circuit breaker can be accurately realized based on this.

[0126] The present specification also provides a computing device, including a memory and a processor, wherein executable codes are stored in the memory, and when the processor executes the executable codes, a method in any embodiment of the present specification is implemented.

[0127] The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute a method in any one of the embodiments in the specification.

[0128] The information interaction, execution process, etc. between the units in the above-mentioned device are based on the same concept as the method embodiments of this specification. For specific contents, please refer to the description in the method embodiments of this specification and will not be repeated here.

[0129] The modules or units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. The above disclosure is only the preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiment are implemented, and the equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A circuit breaker mechanical fault diagnosis method based on acoustic signals, characterized in that: include: Collect the sound signal generated by the circuit breaker to be diagnosed during the opening and closing process; Preprocessing the sound signal to obtain a preprocessed sound signal containing effective information; Performing denoising on the preprocessed sound signal based on an adversarial autoencoder to obtain a denoised sound signal; Extracting time-frequency features of the de-noised sound signal based on short-time Fourier transform to obtain a time-frequency spectrum; Performing zero-crossing rate and kurtosis detection on the denoised sound signal respectively to obtain a zero-crossing rate feature and a kurtosis feature; The mechanical fault of the circuit breaker to be diagnosed is diagnosed according to the time-frequency spectrum diagram, the zero-crossing rate feature and the kurtosis feature.

2. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: The preprocessing of the sound signal comprises: Performing frame processing on the sound signal to obtain a frame signal; wherein there is overlap between adjacent frames during the frame processing; Performing windowing processing on each frame of the framed signal to obtain a frame matrix; Effective feature extraction is performed on the frame matrix to obtain the preprocessed sound signal.

3. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 2, characterized in that: The performing windowing processing on each frame of the framed signal comprises: For each frame signal x k [n], perform windowing processing by the following calculation formula to obtain N windowed signals; in, It is used to represent the nth windowed signal obtained by the kth frame signal, x k [n] is used to represent the kth frame signal of the framed signal, w[n] is used to represent the window function, N is the frame length, 0≤n<N; Convert each frame signal into a row vector and construct the following frame matrix: Wherein, X is used to represent the frame matrix, and K is used to represent the number of frames.

4. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 3, characterized in that: The effective feature extraction for the frame matrix includes: For each frame signal, the total energy value of the corresponding frame signal is calculated by the following calculation formula: Among them, E k Used to characterize the total energy value corresponding to the k-th frame signal, Used to represent the nth windowed signal corresponding to the kth frame signal; Each frame signal whose total energy value is greater than a preset threshold is determined as the preprocessed sound signal.

5. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: The construction process of the adversarial autoencoder includes: Preprocessing the historical preprocessed sound signal data to obtain preprocessed sample data; wherein the preprocessed sample data conforms to the input format of the model; Mapping the preprocessed sample data to a latent space based on an encoder to generate a low-dimensional latent representation, and remapping the latent representation to an original data space based on a decoder to obtain first sample training data without noise; Performing autoencoder training; specifically, the steps include: mapping the noise preprocessed sample data to a latent space to obtain a latent representation, and converting the latent representation to the original data space based on a decoder to obtain a second sample training data containing noise; constructing a reconstruction loss based on the difference between the second sample training data containing noise and the first sample training data containing no noise, and performing training by minimizing the reconstruction loss; Conduct adversarial training; specifically, the method includes: selecting a suitable prior distribution as the target distribution of the latent space; inputting the preprocessed sample data as real data into the discriminator, and generating a potential representation through the encoder using the preprocessed sample data with added noise, and inputting the potential representation into the discriminator as forged data; optimizing the discriminator through the cross entropy loss, so that the potential representation generated by the encoder is as close to the prior distribution as possible; The autoencoder training and the adversarial training are alternately iterated to obtain an adversarial autoencoder model for denoising the preprocessed sound signal.

6. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: The extracting time-frequency features of the denoised sound signal based on short-time Fourier transform includes: At each time point t on the time axis, a window function is used to intercept the local signal; Perform Fourier transform on each intercepted local signal to obtain the frequency characteristics of each local signal; The time-frequency spectrum is constructed by using the frequency characteristics of the local signal intercepted at each time point on the time axis.

7. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 6, characterized in that: The extracting time-frequency features of the denoised sound signal based on short-time Fourier transform includes: For a given signal x(t), the following calculation formula is used to extract time-frequency features: Among them, X(t,f) is used to characterize the frequency characteristics corresponding to the time point t, w(τ-t) is used to characterize the window function, and f is the frequency.

8. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: Performing kurtosis detection on the de-noised sound signal includes: For each frame signal in the preprocessed sound signal, the kurtosis is calculated using the following calculation formula: Among them, Kurtosis is used to characterize the kurtosis characteristics, x i is the i-th signal in the current frame signal of the preprocessed sound signal after denoising, n is the number of data points, μ is the mean of the data, and σ is the variance.

9. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: Performing zero-crossing rate detection on the de-noised sound signal includes: For each frame signal in the preprocessed sound signal, the zero-crossing rate is calculated using the following calculation formula: Among them, ZCR is used to represent the zero-crossing rate, N is used to represent the frame length, and x[n] is used to represent the nth signal of the current frame signal in the denoised preprocessed signal.

10. The circuit breaker mechanical fault diagnosis method based on acoustic signals according to claim 1, characterized in that: The diagnosing the mechanical fault of the circuit breaker to be diagnosed according to the time-frequency spectrum diagram, the zero-crossing rate feature and the kurtosis feature comprises: Inputting the one-dimensional zero-crossing rate feature and the kurtosis feature into a one-dimensional convolutional neural network of a pre-trained classification model to obtain a one-dimensional first feature vector; Inputting the two-dimensional time-frequency spectrum graph into a two-dimensional convolutional neural network of a pre-trained classification model to obtain a two-dimensional second feature vector; The first eigenvector and the second eigenvector are concatenated and input into a fully connected layer to obtain a mechanical fault diagnosis result of the circuit breaker.

Images