CN115910097A - Audible signal identification method and system for latent fault of high-voltage circuit breaker - Google Patents

Abstract

The invention discloses a method and a system for recognizing audible signals of latent faults of a high-voltage circuit breaker. Wherein, the method comprises the following steps: the method comprises the following steps: converting a breaker action sound signal collected by a microphone into a two-dimensional time frequency spectrum with two dimensions of a time domain and a frequency domain; performing feature extraction and dimension reduction on the two-dimensional time spectrum through a Mel cepstrum coefficient, a gamma-pass filtering cepstrum coefficient and a power law normalization cepstrum coefficient, determining a Mel cepstrum feature, a gamma-pass filtering cepstrum feature and a power law normalization cepstrum feature, and forming a cepstrum feature matrix according to the Mel cepstrum feature, the gamma-pass filtering cepstrum feature and the power law normalization cepstrum feature; and taking the convolutional neural network as a classifier, and identifying the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix. The accuracy is improved, and meanwhile, the calculation speed is not obviously reduced, so that the calculation efficiency is optimized.

Description

A method and system for identifying audible signals of latent faults of high-voltage circuit breakers

Technical Field

The present invention belongs to the technical field of abnormal identification of high-voltage electrical equipment, and more specifically, to a method and system for identifying audible sound signals of latent faults of high-voltage circuit breakers.

Background Art

As an important power transmission and transformation equipment, the correct operation of the operating mechanism of high-voltage circuit breakers is directly related to the safe operation of the system. According to statistics, mechanical failures of the operating mechanism of high-voltage circuit breakers account for 70% to 80% of all failures. It is of great significance to carry out mechanical state assessment and fault diagnosis technology and detection equipment for high-voltage circuit breaker operating mechanisms.

Most of the common faults in the actual application of high-voltage circuit breakers (refusal to operate, malfunction, jamming and fracture, etc.) are closely related to their operating mechanisms. Such faults are often caused by the gradual development and accumulation of latent mechanical faults (oil buffer leakage, spring fatigue, transmission pin wear, etc.). The early characteristics of latent faults are not obvious enough. If they are not discovered and handled in time, they are easy to develop into serious faults and cause more serious economic losses. The current research on latent faults mainly focuses on overheating and discharge, while there are fewer studies on mechanical latent faults. Therefore, further research is needed on their feature extraction and early identification.

The identification of mechanical faults of circuit breakers mainly focuses on the analysis and processing of vibration or soundprint signals during operation. Vibration and soundprint are different manifestations of mechanical waves in different media. However, in actual use, vibration sensors need to be deployed by drilling holes on the circuit breaker, and different deployment positions have a strong impact on the monitoring results. Soundprint sensors do not need to be in direct contact with the circuit breaker body, and a small deviation in the measurement point position will result in a small difference in the response to the sound signal, making them more convenient in field applications.

Summary of the invention

According to the present invention, a method and system for identifying audible sound signals of latent faults of high-voltage circuit breakers are provided to solve the technical problem that current research on latent faults mainly focuses on overheating and discharge, while research on mechanical latent faults is relatively less.

According to a first aspect of the present invention, there is provided a method for identifying audible sound signals of latent faults of a high-voltage circuit breaker, comprising:

The circuit breaker action sound signal collected by the microphone is converted into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain;

Through the Mel cepstral coefficients, the gamma-tone filter cepstral coefficients and the power-law normalized cepstral coefficients, the two-dimensional time-frequency spectrum is subjected to feature extraction and dimension reduction, the Mel cepstral features, the gamma-tone filter cepstral features and the power-law normalized cepstral features are determined, and a cepstral feature matrix is constructed according to the Mel cepstral features, the gamma-tone filter cepstral features and the power-law normalized cepstral features;

The convolutional neural network is used as a classifier, and the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker is identified according to the cepstrum feature matrix.

Optionally, the circuit breaker action sound signal collected by the microphone is converted into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain, including:

The Hamming window is determined according to the following formula:

Where w(n) is the Hamming window and N is the length;

Perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

且k≤N-1

And k≤N-1

Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through Mel-cepstral coefficients to determine Mel-cepstral features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Pass the power spectrum through a Mel-scale filter bank with a triangular frequency domain to determine Mel-cepstrum parameters;

Taking the natural logarithm of the Mel-cepstrum parameter and obtaining an M-dimensional Mel-cepstrum through discrete cosine transform, and performing mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature;

Wherein, M is the number of Mel filters.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through gammatone filter cepstrum coefficients to determine gammatone filter cepstrum features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to determine gammatone filter cepstrum parameters;

The gammatone filter cepstrum parameters are subjected to natural logarithm and then discrete cosine transform to obtain an M-dimensional gammatone filter cepstrum, and the gammatone filter cepstrum is mean-normalized to determine gammatone filter cepstrum features.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through power-law normalized cepstrum coefficients to determine power-law normalized cepstrum features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Passing the power spectrum through a Gammatone filter bank consisting of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter;

According to the power spectrum after the filter, each frame and the two frames before and after are smoothed, the low-frequency part is filtered out, asymmetric noise suppression is performed, and the average power in the middle is calculated;

Determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power;

Normalizing the original short-time energy spectrum in the time-frequency domain according to the time-average and frequency-average transfer functions to determine the energy spectrum after time-frequency normalization;

The energy spectrum after time-frequency normalization is subjected to average power normalization according to the average power estimation value, and discrete cosine transformation and mean normalization are performed to determine power-law normalized cepstrum features.

Optionally, a convolutional neural network is used as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix, including:

Using the cepstral feature matrix as an input layer, a hybrid cepstral coefficient-convolutional neural network recognition model is constructed;

According to the mixed cepstral coefficient-convolutional neural network recognition model, the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker is identified.

According to another aspect of the present invention, there is also provided a system for identifying latent fault audible signals of a high-voltage circuit breaker, comprising:

An action signal conversion module is used to convert the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain;

A feature matrix module is constructed, which is used to extract features and reduce the dimension of the two-dimensional time-frequency spectrum through Mel cepstral coefficients, gamma-tone filter cepstral coefficients and power-law normalized cepstral coefficients, determine Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features, and construct a cepstral feature matrix according to the Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features;

The fault signal identification module is used to use the convolutional neural network as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix.

Optionally, the conversion action signal module includes:

The Hamming window determination submodule is used to determine the Hamming window according to the following formula:

Where w(n) is the Hamming window and N is the length;

The spectrum matrix determination submodule is used to perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

且k≤N-1

And k≤N-1

Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A Mel-cepstrum parameter determination submodule is used to determine the Mel-cepstrum parameters by passing the power spectrum through a Mel-scale filter bank with a triangular frequency domain;

A Mel-cepstrum feature determination submodule is used to obtain an M-dimensional Mel-cepstrum by discrete cosine transform after taking the natural logarithm of the Mel-cepstrum parameter, and to perform mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature;

Wherein, M is the number of Mel filters.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A gammatone filter cepstrum parameter determination submodule is used to determine the gammatone filter cepstrum parameters by passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters;

A gammatone filter cepstrum feature submodule is determined, which is used to obtain an M-dimensional gammatone filter cepstrum by discrete cosine transform after taking the natural logarithm of the gammatone filter cepstrum parameters, and to perform mean normalization on the gammatone filter cepstrum to determine the gammatone filter cepstrum feature.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A filtering power spectrum submodule is obtained, which is used to pass the power spectrum through a Gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter;

The submodule for calculating the average power in the middle time is used to smooth each frame and the two frames before and after it according to the power spectrum after passing through the filter, filter out the low-frequency part, perform asymmetric noise suppression, and calculate the average power in the middle time;

A submodule for determining an average transfer function, used for determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power;

A submodule for determining the energy spectrum after time-frequency normalization is used to normalize the original short-time energy spectrum in the time-frequency domain according to the time average and frequency average transfer functions to determine the energy spectrum after time-frequency normalization;

The power-law normalized cepstrum feature submodule is used to perform average power normalization on the energy spectrum after time-frequency normalization according to the average power estimation value, perform discrete cosine transform and mean normalization, and determine the power-law normalized cepstrum feature.

Optionally, the fault signal identification module includes:

Constructing a neural network recognition model submodule, which is used to use the cepstral feature matrix as an input layer to construct a mixed cepstral coefficient-convolutional neural network recognition model;

The fault type identification submodule is used to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the mixed cepstral coefficient-convolutional neural network identification model.

Therefore, starting from the needs of power grid equipment operation and inspection, combined with the company's professional characteristics of operation and inspection and the development trend of building an intelligent operation and inspection system, the diagnostic method is studied with the potential mechanical fault soundprint of the circuit breaker that is more in line with the on-site application scenario as the object. Through the integration of artificial intelligence technology and traditional operation and inspection business, the intelligent identification of abnormal working conditions of the operating mechanism of the high-voltage circuit breaker is realized. Effectively improve the equipment status control and operation and inspection management penetration, realize data-driven operation and inspection business innovation and efficiency improvement, and comprehensively promote the innovation of operation and inspection work methods and production management models. While improving the accuracy, the calculation speed has not decreased significantly, thereby optimizing the calculation efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

A more complete understanding of exemplary embodiments of the present invention may be obtained by referring to the following drawings:

FIG1 is a schematic flow chart of a method for identifying audible sound signals of latent faults of a high-voltage circuit breaker according to the present embodiment;

FIG2 is a schematic diagram of an audible signal recognition system for a latent fault of a high-voltage circuit breaker according to the present embodiment.

DETAILED DESCRIPTION

Now, exemplary embodiments of the present invention are described with reference to the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to disclose the present invention in detail and completely and to fully convey the scope of the present invention to those skilled in the art. The terms used in the exemplary embodiments shown in the accompanying drawings are not intended to limit the present invention. In the accompanying drawings, the same units/elements are marked with the same reference numerals.

Unless otherwise specified, the terms (including technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is understood that the terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

According to a first aspect of the present invention, a method 100 for identifying an audible signal of a latent fault of a high-voltage circuit breaker is provided. Referring to FIG. 1 , the method 100 comprises:

S101: converting the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain;

S102: performing feature extraction and dimensionality reduction on the two-dimensional time-frequency spectrum through Mel cepstral coefficients, gamma-tone filter cepstral coefficients and power-law normalized cepstral coefficients, determining Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features, and constructing a cepstral feature matrix according to the Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features;

S103: Using the convolutional neural network as a classifier, and identifying the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix.

Specifically, the present invention collects the soundprint signals of the circuit breaker under different latent fault operating states by building a circuit breaker test platform, and uses a mixed cepstral calculation method and a CNN convolutional neural network structure to distinguish the sound signals of the high-voltage circuit breaker opening and closing under different fault states. First, a sound signal sample library is formed by on-site collection; then, the Mel Frequency Cepstral Coefficient (MFCC), Gammatone Filter Cepstral Coefficient (GFCC), and Power-Normalized Cepstral Coefficient (PNCC) are used to reduce the dimension and extract preliminary features of the original signal; finally, a convolutional neural network (CNN) is introduced as a classifier, and the mixed cepstral (Mixed-Cepstral Coefficient) formed by the triple features of MFCC, GFCC, and PNCC is used to form a classification model for six operating states, and the effectiveness of the model is verified by the data set.

The circuit breaker acoustic signal recognition first converts the circuit breaker action acoustic signal collected by the microphone into a time-frequency spectrum with two dimensions in the time domain and frequency domain. Then, the original time-frequency spectrum is feature extracted and dimension reduced by calculating three types of inverse spectrums, such as MFCCs, GFCCs and PNCCs. Finally, a convolutional neural network is used as a classifier to identify the fault type. The entire process can be roughly divided into three parts: acoustic signal acquisition, preprocessing and pattern recognition. Among them, the preprocessing method of the sound signal is the most important. Its main function is to extract features and compress data of the original time domain signal of the circuit breaker, thereby reducing the amount of calculation of the subsequent recognition model and improving the recognition effect.

1. Calculation of frequency spectrum of circuit breaker action sound signal

The time domain signal of the circuit breaker sound signal is generally a one-dimensional pulse signal, and its characteristic information is not obvious enough. It can be converted into a two-dimensional time-frequency spectrum by using the short-time discrete Fourier transform method, which is conducive to improving the recognition speed and recognition accuracy of the deep learning model. In the process of short-time Fourier transform, frame division, windowing and discrete Fourier transform are required. Among them, the window function can usually select the Hamming window to reduce the spectrum leakage caused by the Fourier transform. The formula of the Hamming window w(n) with a length of N is as follows:

The time-frequency spectrum matrix is obtained by performing short-time discrete Fourier transform on the discretized time domain frame:

Among them, k is the frequency point number, and x(n) is the original discrete time domain signal.

2 Quasi-steady-state process analysis

The characteristics of latent circuit breaker faults are generally not obvious. Therefore, when conducting circuit breaker voiceprint diagnosis, its voiceprint features must be extracted while ensuring the speed of sound signal recognition, thereby improving the recognition accuracy. The cepstral coefficient calculation method widely used in the field of speech recognition can compress the sample data while retaining the key voiceprint information of closing, thereby realizing the compression and feature extraction of circuit breaker sound signals, which helps to improve the diagnosis speed and accuracy of subsequent convolutional neural network classifiers.

Among all kinds of cepstrum coefficients, MFCC is based on the auditory model, GFCC is based on the eardrum model, and PNCC has more advantages in extracting sound features under a noisy background. The above cepstrum coefficients have been applied to a certain extent in the field of speech recognition. Therefore, this study selects MFCC, GFCC, and PNCC as basic cepstrum features to form a cepstrum feature matrix for subsequent feature fusion and sound signal recognition.

(1) MFCC calculation

MFCC is a cepstrum parameter based on the human hearing perception characteristics. In the frequency domain, the height of the sound heard by the human ear is not linearly related to the frequency, but in the Mel domain, the human ear perception is proportional to the Mel frequency. The relationship can be expressed as follows:

Mel(f)＝2595lg(1+f/700) (3)

The calculation of Mel frequency cepstral coefficients is performed in frames. The following are the specific calculation steps of Mel frequency cepstral coefficients:

First, the spectrum X(k) of each frame is calculated according to formula (1), and the power spectrum is obtained by taking the modulus and squaring each frame X(k). The power spectrum is passed through a Mel-scale filter bank with a triangular frequency domain to obtain a new parameter R(k). The lower limit of the filter bank frequency is f _min , the upper limit is f _max , and M is the number of Mel filters.

Then take the natural logarithm of R(k):

Then the M-dimensional MFCC is obtained through discrete cosine transform (DCT):

Finally, mean normalization is required. After the above steps, M-dimensional MFCC can be obtained.

(2) GFCC calculation

The extraction process of GFCC is almost the same as that of MFCC. The difference between the two is that the filter bank through which the power spectrum passes is a Gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters, rather than a Mel-scale filter. The upper and lower limits of the filter bank frequency are also f _max and f _min , and the subsequent calculation steps are also the same.

(3) PNCC calculation

The first two steps of the PNCC extraction process are the same as those of the GFCC. When the power spectrum passes through the Gammatone filter, P[m,l] is obtained, where l represents the channel number. Each frame is smoothed with the two frames before and after, and the average power is calculated:

然后对不同信道进行平滑处理:The obtained mid-time average power is used for the subsequent environmental background noise estimation and compensation. The spectral subtraction method is used to filter out the low-frequency part to achieve the purpose of noise suppression, that is, asymmetric noise suppression is performed to obtain

Then smooth the different channels:

l ₁ =max(lp,1) (8)

l ₂ =min(l+p,M) (9)

Where M represents the number of channels, and p is generally set to 4.

对P[m,l]时-频域归一化：use

Normalize P[m,l] in the time-frequency domain:

The average power estimate μ[m] can be used to normalize T[m,l] to the average power:

Where k is a coefficient and can be set to any constant.

In order to be closer to the compressed sensing characteristics of the human auditory nerve, PNCC uses power-law nonlinear compression, which is different from the logarithmic nonlinearity used by MFCC:

V[m,l]＝U[m,l] ^1/15 (13)

Finally, discrete cosine transform and mean normalization are performed to obtain PNCC.

After completing the calculation of MFCC, GFCC, and PNCC respectively, the three are merged into a [Z×M×3] cepstrum feature matrix, where Z is the time domain component, which depends on the number of time frames of the original time-frequency spectrum; M is the frequency domain component, which is equal to the number of filters used in the calculation of each cepstrum coefficient. The more the number, the richer the information, but the larger the amount of data. The general setting range is 40 to 48.

5.3 Convolutional Neural Network Calculation

Since the cepstral feature matrix is a three-dimensional map composed of multiple two-dimensional maps of the same size, the convolutional neural network (CNN), which is representative in the field of image recognition, can be introduced as a classifier for the cepstral feature matrix of acoustic signals. In the field of image recognition, CNN usually splits color images into three color layers of red, green and blue (RGB) as the input layer of the network, so as to learn and perceive the characteristics of different color changes. Similarly, this study uses the cepstral feature matrix composed of three cepstrals of the acoustic signal as the input layer, and constructs a mixed cepstral coefficient-convolutional neural network (MCC-CNN) recognition model to perform sound classification and recognition. Compared with the artificially designed cepstral mixing method, the fusion of the three cepstrals through the learning mechanism of the deep neural network can make the fusion method of the mixed cepstral adaptive.

Since the circuit breaker closing voiceprint data has been compressed, it can be realized through a VGG-like lightweight CNN network. The network as a whole contains 3 convolution-pooling layers and 4 fully connected layers. In the network, Dropout and batch normalization need to be added to prevent overfitting and gradient disappearance. The detailed structure is shown in Table 1.

Table 1 VGG-like lightweight CNN network structure

Tab, 1 VGG-like lightweight CNN Structure

In order to verify the effectiveness of the recognition model of the method, the recognition success rate and operation time of MCC-CNN, MFCC-CNN, GFCC-CNN, PNCC-CNN and conventional CNN models were compared. The results are shown in Table 2. It can be seen that the MCC-CNN model performs best in recognition success rate, which proves the superiority of the MCC-CNN sound recognition model proposed in the present invention.

Compared with the method of directly using the time-frequency spectrum graph to input the CNN network for training and recognition, the amount of data and calculation of MCC-CNN has been greatly reduced after preprocessing with the mixed cepstrum calculation. The reduction in sample data also means that the recognition difficulty of the deep neural network is reduced, so the recognition rate can be improved and the recognition time can be reduced. Compared with the method of using a single cepstrum coefficient preprocessing method, MCC-CNN uses more types of cepstrum features to adapt to a variety of latent mechanical fault sound signals. Moreover, from the results, the calculation speed of the trained MCC-CNN is not significantly inferior to the single cepstrum coefficient method with smaller data volume. The accuracy is improved without significantly reducing the calculation speed. This is because the Dropout operation eliminates many redundant neuron connections, thereby optimizing the calculation efficiency.

Table 2 Comparison of the effects of different pre-processing methods

Tab.5 Comparison of different preprocessing methods

Optionally, the circuit breaker action sound signal collected by the microphone is converted into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain, including:

The Hamming window is determined according to the following formula:

Where w(n) is the Hamming window and N is the length;

Perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

且k≤N-1

And k≤N-1

Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through Mel-cepstral coefficients to determine Mel-cepstral features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Pass the power spectrum through a Mel-scale filter bank with a triangular frequency domain to determine Mel-cepstrum parameters;

Taking the natural logarithm of the Mel-cepstrum parameter and obtaining an M-dimensional Mel-cepstrum through discrete cosine transform, and performing mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature;

Wherein, M is the number of Mel filters.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through gammatone filter cepstrum coefficients to determine gammatone filter cepstrum features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to determine gammatone filter cepstrum parameters;

The gammatone filter cepstrum parameters are subjected to natural logarithm and then discrete cosine transform to obtain an M-dimensional gammatone filter cepstrum, and the gammatone filter cepstrum is mean-normalized to determine gammatone filter cepstrum features.

Optionally, feature extraction and dimension reduction are performed on the two-dimensional time-frequency spectrum through power-law normalized cepstrum coefficients to determine power-law normalized cepstrum features, including:

Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum;

Passing the power spectrum through a Gammatone filter bank consisting of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter;

According to the power spectrum after the filter, each frame and the two frames before and after are smoothed, the low-frequency part is filtered out, asymmetric noise suppression is performed, and the average power in the middle is calculated;

Determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power;

Normalizing the original short-time energy spectrum in the time-frequency domain according to the time-average and frequency-average transfer functions to determine the energy spectrum after time-frequency normalization;

The energy spectrum after time-frequency normalization is subjected to average power normalization according to the average power estimation value, and discrete cosine transformation and mean normalization are performed to determine power-law normalized cepstrum features.

Optionally, a convolutional neural network is used as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix, including:

Using the cepstral feature matrix as an input layer, a hybrid cepstral coefficient-convolutional neural network recognition model is constructed;

According to the mixed cepstral coefficient-convolutional neural network recognition model, the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker is identified.

According to another aspect of the present invention, there is also provided a system for identifying latent fault audible signals of a high-voltage circuit breaker, comprising:

An action signal conversion module is used to convert the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain;

A feature matrix module is constructed, which is used to extract features and reduce the dimension of the two-dimensional time-frequency spectrum through Mel cepstral coefficients, gamma-tone filter cepstral coefficients and power-law normalized cepstral coefficients, determine Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features, and construct a cepstral feature matrix according to the Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features;

The fault signal identification module is used to use the convolutional neural network as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix.

Optionally, the conversion action signal module includes:

The Hamming window determination submodule is used to determine the Hamming window according to the following formula:

Where w(n) is the Hamming window and N is the length;

The spectrum matrix determination submodule is used to perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

且k≤N-1

And k≤N-1

Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A Mel-cepstrum parameter determination submodule is used to determine the Mel-cepstrum parameters by passing the power spectrum through a Mel-scale filter bank with a triangular frequency domain;

A Mel-cepstrum feature determination submodule is used to obtain an M-dimensional Mel-cepstrum by discrete cosine transform after taking the natural logarithm of the Mel-cepstrum parameter, and to perform mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature;

Wherein, M is the number of Mel filters.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A gammatone filter cepstrum parameter determination submodule is used to determine the gammatone filter cepstrum parameters by passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters;

A gammatone filter cepstrum feature submodule is determined, which is used to obtain an M-dimensional gammatone filter cepstrum by discrete cosine transform after taking the natural logarithm of the gammatone filter cepstrum parameters, and to perform mean normalization on the gammatone filter cepstrum to determine the gammatone filter cepstrum feature.

Optionally, a feature matrix module is constructed, including:

A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum;

A filtering power spectrum submodule is obtained, which is used to pass the power spectrum through a Gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter;

The submodule for calculating the average power in the middle time is used to smooth each frame and the two frames before and after it according to the power spectrum after passing through the filter, filter out the low-frequency part, perform asymmetric noise suppression, and calculate the average power in the middle time;

A submodule for determining an average transfer function, used for determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power;

A submodule for determining the energy spectrum after time-frequency normalization is used to normalize the original short-time energy spectrum in the time-frequency domain according to the time average and frequency average transfer functions to determine the energy spectrum after time-frequency normalization;

The power-law normalized cepstrum feature submodule is used to perform average power normalization on the energy spectrum after time-frequency normalization according to the average power estimation value, perform discrete cosine transform and mean normalization, and determine the power-law normalized cepstrum feature.

Optionally, the fault signal identification module includes:

Constructing a neural network recognition model submodule, which is used to use the cepstral feature matrix as an input layer to construct a mixed cepstral coefficient-convolutional neural network recognition model;

The fault type identification submodule is used to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the mixed cepstral coefficient-convolutional neural network identification model.

A DC voltage robust control system 500 when a VSC is connected to a weak power grid according to an embodiment of the present invention corresponds to a DC voltage robust control method 300 when a VSC is connected to a weak power grid according to another embodiment of the present invention, and will not be described in detail here.

Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

The present application is described with reference to the flowchart and/or block diagram of the method, device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and/or box in the flowchart and/or block diagram, and the combination of the process and/or box in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one process or multiple processes in the flowchart and/or one box or multiple boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims (12)

1. A method for identifying audible sound signals of latent faults of high-voltage circuit breakers, comprising: Converting the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain; Through the Mel cepstral coefficients, the gamma-tone filter cepstral coefficients and the power-law normalized cepstral coefficients, the two-dimensional time-frequency spectrum is subjected to feature extraction and dimension reduction, the Mel cepstral features, the gamma-tone filter cepstral features and the power-law normalized cepstral features are determined, and a cepstral feature matrix is constructed according to the Mel cepstral features, the gamma-tone filter cepstral features and the power-law normalized cepstral features; The convolutional neural network is used as a classifier, and the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker is identified according to the cepstrum feature matrix. 2. The method according to claim 1 is characterized in that converting the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain comprises: The Hamming window is determined according to the following formula:

Where w(n) is the Hamming window and N is the length; Perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

And k≤N-1 Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal. 3. The method according to claim 1, characterized in that the mel-cepstrum coefficients are used to extract features and reduce the dimension of the two-dimensional time-frequency spectrum to determine the mel-cepstrum features, comprising: Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum; Pass the power spectrum through a Mel-scale filter bank with a triangular frequency domain to determine Mel-cepstrum parameters; Taking the natural logarithm of the Mel-cepstrum parameter and obtaining an M-dimensional Mel-cepstrum through discrete cosine transform, and performing mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature; Wherein, M is the number of Mel filters. 4. The method according to claim 1, characterized in that the feature extraction and dimension reduction of the two-dimensional time-frequency spectrum are performed through gammatone filter cepstrum coefficients to determine the gammatone filter cepstrum features, comprising: Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum; Passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to determine gammatone filter cepstrum parameters; The gammatone filter cepstrum parameters are subjected to natural logarithm and then discrete cosine transform to obtain an M-dimensional gammatone filter cepstrum, and the gammatone filter cepstrum is mean-normalized to determine gammatone filter cepstrum features. 5. The method according to claim 1, characterized in that the power-law normalized cepstrum coefficients are used to extract features and reduce the dimension of the two-dimensional time-frequency spectrum to determine the power-law normalized cepstrum features, comprising: Taking the modulus of each frame of the two-dimensional time-frequency spectrum and squaring it to obtain a power spectrum; Passing the power spectrum through a Gammatone filter bank consisting of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter; According to the power spectrum after the filter, each frame and the two frames before and after are smoothed, the low-frequency part is filtered out, asymmetric noise suppression is performed, and the average power in the middle is calculated; Determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power; Normalizing the original short-time energy spectrum in the time-frequency domain according to the time-average and frequency-average transfer functions to determine the energy spectrum after time-frequency normalization; The energy spectrum after time-frequency normalization is subjected to average power normalization according to the average power estimation value, and discrete cosine transformation and mean normalization are performed to determine power-law normalized cepstrum features. 6. The method according to claim 1 is characterized in that a convolutional neural network is used as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix, comprising: Using the cepstral feature matrix as an input layer, a hybrid cepstral coefficient-convolutional neural network recognition model is constructed; According to the mixed cepstral coefficient-convolutional neural network recognition model, the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker is identified. 7. A system for identifying latent faults of high-voltage circuit breakers using audible signals, comprising: An action signal conversion module is used to convert the circuit breaker action sound signal collected by the microphone into a two-dimensional time-frequency spectrum having two dimensions of time domain and frequency domain; A feature matrix module is constructed, which is used to extract features and reduce the dimension of the two-dimensional time-frequency spectrum through Mel cepstral coefficients, gamma-tone filter cepstral coefficients and power-law normalized cepstral coefficients, determine Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features, and construct a cepstral feature matrix according to the Mel cepstral features, gamma-tone filter cepstral features and power-law normalized cepstral features; The fault signal identification module is used to use the convolutional neural network as a classifier to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the cepstrum feature matrix. 8. The system according to claim 7, characterized in that the conversion action signal module comprises: The Hamming window determination submodule is used to determine the Hamming window according to the following formula:

Where w(n) is the Hamming window and N is the length; The spectrum matrix determination submodule is used to perform short-time discrete Fourier transform on the circuit breaker action sound signal to determine the two-dimensional time-frequency spectrum matrix:

And k≤N-1 Among them, X(k) is the frequency spectrum of each frame of the two-dimensional time-frequency spectrum matrix, k is the frequency point number, and x(n) is the circuit breaker action sound signal. 9. The system according to claim 7, characterized in that the feature matrix module comprises: A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum; A Mel-cepstrum parameter determination submodule is used to determine the Mel-cepstrum parameters by passing the power spectrum through a Mel-scale filter bank with a triangular frequency domain; A Mel-cepstrum feature determination submodule is used to obtain an M-dimensional Mel-cepstrum by discrete cosine transform after taking the natural logarithm of the Mel-cepstrum parameter, and to perform mean normalization on the M-dimensional Mel-cepstrum to determine the Mel-cepstrum feature; Wherein, M is the number of Mel filters. 10. The system according to claim 7, characterized in that the feature matrix module comprises: A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum; A gammatone filter cepstrum parameter determination submodule is used to determine the gammatone filter cepstrum parameters by passing the power spectrum through a gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters; A gammatone filter cepstrum feature submodule is determined, which is used to obtain an M-dimensional gammatone filter cepstrum by discrete cosine transform after taking the natural logarithm of the gammatone filter cepstrum parameters, and to perform mean normalization on the gammatone filter cepstrum to determine the gammatone filter cepstrum feature. 11. The system according to claim 7, characterized in that the feature matrix module comprises: A power spectrum submodule is used to obtain a power spectrum by taking the modulus and then squaring each frame of the two-dimensional time spectrum; A filtering power spectrum submodule is obtained, which is used to pass the power spectrum through a Gammatone filter bank composed of M gamma filters with different scale parameters and shape parameters to obtain a power spectrum after passing through the filter; The submodule for calculating the average power in the middle time is used to smooth each frame and the two frames before and after it according to the power spectrum after passing through the filter, filter out the low-frequency part, perform asymmetric noise suppression, and calculate the average power in the middle time; A submodule for determining an average transfer function, used for determining the time-averaged and frequency-averaged transfer functions according to the intermediate time average power; A submodule for determining the energy spectrum after time-frequency normalization is used to normalize the original short-time energy spectrum in the time-frequency domain according to the time average and frequency average transfer functions to determine the energy spectrum after time-frequency normalization; The power-law normalized cepstrum feature submodule is used to perform average power normalization on the energy spectrum after time-frequency normalization according to the average power estimation value, perform discrete cosine transform and mean normalization, and determine the power-law normalized cepstrum feature. 12. The system according to claim 7, characterized in that the fault signal identification module comprises: Constructing a neural network recognition model submodule, which is used to use the cepstral feature matrix as an input layer to construct a mixed cepstral coefficient-convolutional neural network recognition model; The fault type identification submodule is used to identify the fault type of the audible sound signal of the latent fault of the high-voltage circuit breaker according to the mixed cepstral coefficient-convolutional neural network identification model.

Images