Switch cabinet partial discharge identification method
By preprocessing and feature fusion of the ultrasonic signals from the switchgear, combined with a bidirectional recurrent neural network model, the problem of inaccurate partial discharge identification in existing technologies has been solved, achieving more efficient partial discharge type identification.
Patent Information
- Application Number
- CN202511526858.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-24
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies for partial discharge identification in switchgear suffer from nonlinear signal analysis, insufficient ultrasonic database data, complex feature parameter processing, inadequate LPCC and MFCC capabilities, and the inability of the GRU model to process long-sequence signals, resulting in poor identification performance.
By preprocessing the ultrasound signal to convert it into a speech-like signal, extracting features using LPCC and MFCC, fusing Fisher's criterion to filter feature parameters, and employing a bidirectional recurrent neural network model for recognition, including bidirectional GRU layers and fully connected layers, the network structure and parameter settings are optimized to improve the recognition rate.
It improves the resolution and short-term stability of partial discharge signals, enhances the self-learning ability of the identification model, and achieves efficient identification of partial discharge types.
Smart Images
Figure CN121522374A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of signal and discharge identification, and particularly to a switch cabinet partial discharge identification method. BACKGROUND
[0002] When the switch cabinet has partial discharge phenomenon, the field strength will change, which will cause different types of discharge modes such as surface discharge, internal discharge, corona discharge, etc. The current partial discharge defect signal used for analysis does not present linear output signal as a whole, the analysis angle of the defect is less, the partial discharge ultrasonic database data is less, and the ultrasonic feature parameter processing and analysis is more complex; the LPCC has weak ability to express frequency profile information of the partial discharge sound signal, the MFCC has weak ability to express time domain, amplitude, phase and other information of the partial discharge sound signal; the GRU model cannot contact the feature information of the previous and subsequent time, and the GRU has weak ability to process the partial discharge sound signal with long feature sequence. SUMMARY
[0003] In view of the above defects of the prior art, the present application provides a switch cabinet partial discharge identification method, comprising the following steps:
[0004] S1, obtaining an ultrasonic signal of a switch cabinet, and preprocessing the ultrasonic signal;
[0005] S2, converting the preprocessed ultrasonic signal into a speech-like signal, extracting features of the speech-like signal, and generating an input of an identification model;
[0006] S3, taking the extracted features as the input, and obtaining a partial discharge type through the identification model.
[0007] Further, the preprocessing process of step S1 includes: after frame division and windowing processing of the ultrasonic signal, the ultrasonic signal is converted into a speech-like signal by using the relationship between frequency and mel frequency. The speech-like signal can represent the characteristics of a signal that can change at any time within a very short period of time, and can well describe the characteristics of the partial discharge ultrasonic signal, which greatly supports the identification of the model.
[0008] Further, step S2 includes:
[0009] extracting the frequency profile, harmonics and formants of the sound signal through LPCC;
[0010] extracting the pitch information of the sound signal through MFCC;
[0011] fusing the features carried by LPCC and MFCC through Fisher criterion, extracting the more significant LPCC-MFCC features of acoustic features, and constituting a feature sequence.
[0012] Further, the Fisher criterion is:
[0013] ;
[0014] ;
[0015] ;
[0016] wherein, is the between-class scatter; is the within-class scatter; is the Fisher ratio; is the mean of the i-th class signal in the k-th feature component; is the sample number of the i-th class signal; is the k-th feature parameter of the i-th class signal.
[0017] Further, the recognition model is a bidirectional recurrent neural network model.
[0018] Further, the bidirectional recurrent neural network model comprises two layers of gated recurrent units as two hidden layers respectively.
[0019] Compared with the prior art, the present application has the following technical effects:
[0020] The present application converts the ultrasonic signal into a speech-like signal by using the Mel frequency, i.e. stretching the low-frequency part of the ultrasonic signal and compressing the high-frequency component of the ultrasonic signal. This conversion can amplify the low-frequency features valuable for diagnosis, and further improve the resolution of the preprocessed ultrasonic signal, better reflecting the short-time stationarity of the ultrasonic signal. By extracting signal features through linear prediction cepstrum coefficient (LPCC) and Mel frequency cepstrum coefficient (MFCC) methods, and based on the different sensitivity characteristics of the two methods in the feature extraction process, the LPCC and MFCC are fused to select better feature parameters for expressing the discharge signal, and a new fused feature parameter sequence is obtained. In the recognition model, the present application uses a bidirectional recurrent neural network to process the long feature sequence, which has the ability to learn by itself and can automatically find the best network weight. Moreover, two layers of hidden layers with fast convergence speed are set to improve the recognition rate.
[0021] The concept, specific structure and technical effects of the present application will be further described below with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS
[0022] Figure 1 is the MFCC extraction flowchart of a specific embodiment of the present application;
[0023] Figure 2is a LPCC extraction flowchart of one embodiment of the present application.
[0024] Figure 3 is a network structure diagram of a bidirectional recurrent neural network of one embodiment of the present application. DETAILED DESCRIPTION
[0025] The present application is herein described, by way of example only, with reference to certain embodiments thereof. It is to be understood that the terminology used herein is for the purpose of describing specific embodiments only and is not intended to be limiting, since the scope of the present application will be limited only by the appended claims. Numerous specific aspects of the application are described herein. However, in an effort to provide a concise description of the instant application, all features of the application will not necessarily be explicitly re-iterated in the following description. It is noted that the following examples and features of the examples can be combined with each other, where not in contradiction and where not already covered by the above summary.
[0026] In specific embodiments, a switch cabinet partial discharge identification method is provided, comprising the following steps:
[0027] S1, obtaining an ultrasonic signal of a switch cabinet, and preprocessing the ultrasonic signal;
[0028] The preprocessing of the ultrasonic signal is to improve the resolution and maintain the spectral smooth continuity; the preprocessing process includes:
[0029] A window function with high frequency resolution and less spectral leakage is selected to perform frame processing on the collected signal, so that the signal is in a short-time stable state within a short time range;
[0030] The window function has a low-frequency characteristic, and the Hamming window function used in this paper has a narrow main lobe width, a narrow effective noise bandwidth, a high resolution, a low sidelobe attenuation rate, and less spectral leakage:
[0031] ;
[0032] The time sequence of the sound signal to be processed is set as x(n), and the i-th frame of sound signal after frame processing and windowing is x i (m), and the frame length is N.
[0033] In order to further improve the recognition rate of the ultrasonic signal, and further improve the resolution of the ultrasonic signal after the pretreatment, in the embodiment, the ultrasonic signal is converted into a speech-like signal by using the Mel-scale Frequency Cepstral Coefficients (MFCC), that is, the low-frequency part of the ultrasonic signal is stretched, and the high-frequency component of the ultrasonic signal is compressed. Through this processing mode, the low-frequency features valuable for diagnosis can be amplified, and the short-time stationarity of the ultrasonic signal can be better reflected. The extraction method of the Mel-scale Frequency Cepstral Coefficients (MFCC) can adopt the common extraction method in the prior art. The processing method of the Mel-scale Frequency Cepstral Coefficients (MFCC) in the embodiment includes the steps of time domain processing, spectrum analysis, Mel filter compression and cepstrum transformation as shown in Figure 1 .
[0034] The MFCC is a coefficient extracted under the Mel scale frequency scale, which can be expressed according to the existing relationship between the Mel frequency and the frequency f as follows:
[0035]
[0036] S2, extracting features in the ultrasonic signal after the pretreatment, and generating a feature sequence;
[0037] The linear predictive cepstral coefficient and the Mel frequency cepstral coefficient of the converted speech-like signal are extracted, and the LPCC and the MFCC are fused based on the Fisher criterion to obtain a new feature parameter sequence. In the embodiment, the Mel frequency cepstral coefficient is the above-mentioned Mel frequency; the method for obtaining the linear predictive cepstral coefficient adopts the method in the prior art, and the extraction process is as shown in Figure 2 .
[0038] After the two feature parameters LPCC and MFCC are extracted respectively, they are denoted as lpcc i (n) and mfcc i (k) respectively, where the subscript i is the frame number, and n and k are the dimensions of the two feature parameters respectively; the two feature parameters LPCC and MFCC are serially fused, and the fused feature parameter sequence is denoted as C lpmfcc , then C lpmfcc ={lpcc i (n), mfcc i (k)}={lpcc i (1), …, lpcc i (n), mfcc i (1), , mfcc i (k)}; the Fisher ratio of each dimension of the fused feature parameter is calculated by using the Fisher criterion, and the dimension feature with a larger Fisher ratio is selected according to the feature dimension requirement to form a new fused feature parameter sequence Clpmfcc1 .
[0039] ;
[0040] ;
[0041] ;
[0042] Among them, J Fisher Fisher's ratio; The inter-class dispersion represents the mean and variance among different feature components; Within-class dispersion represents the mean and variance among different samples of the same feature component. Fisher's ratio; Let be the mean of the k-th dimension feature components of all signal classes; Let be the mean of the k-th dimension feature component of the i-th type signal; The number of samples for class i signals; Let be the characteristic parameters of the i-th type of signal in the k-th dimension.
[0043] In this embodiment, the J Fisher A larger value indicates a greater contribution of that feature dimension to sound signal recognition. In this embodiment, J is selected through comparison. Fisher Larger signals are used as new high-quality features, and the selected feature sequences are used as input data for subsequent steps. For example, given a dataset containing three features [f1, f2, f3], the calculated J... Fisher They are respectively: J Fisher（f1） =6.2; J Fisher(f2) =4.8; J Fisher(f3) =3.3; By comparison, f1 has the largest Fisher ratio, which contributes the most to signal recognition, while f3 has the largest J... Fisher The value is the smallest and the contribution is the lowest. When selecting features, we will prioritize retaining and selecting f1, and discarding f2 and f3.
[0044] S3. The feature sequence is used as input data and input into a bidirectional recurrent neural network recognition model to obtain the types of partial discharge faults.
[0045] like Figure 3As shown, the bidirectional recurrent neural network recognition model in the embodiment includes an input layer, a BiGRU layer 1, a BiGRU layer 2, a full connection layer 1, a full connection layer 2 and an output layer, the BiGRU layer 1 and the BiGRU layer 2 in series constitute a hidden layer, and the parameters and settings contained in the bidirectional gate control recurrent unit are the same. By setting the parameters such as activation function, loss function and learning rate, and using the above processed signal feature data set as the training set to train the network, the corona discharge, surface discharge and internal discharge can be recognized, and finally a better recognition effect is obtained.
[0046] The design of the hidden layer is crucial to the recognition effect. In the embodiment, the number of neurons with the optimal recognition efficiency is selected by designing the number of layers, so that the recognition effect is optimal, and the selection basis is:
[0047] ;
[0048] In the formula, represents the number of neurons, is the number of training set samples, is the number of input layer neurons, is the number of output layer neurons.
[0049] The activation function in the embodiment adopts a linear rectifier function, which is a segmented function. The negative signal has an inhibitory effect, and the non-negative signal is gradually amplified with a certain constant, so that the signal can have good convergence, which is helpful for the smooth output of the result. The activation function is:
[0050] ;
[0051] In the process of neural network training, the learning rate should not be too large or too small. It is related to whether the result can converge and the convergence speed. Therefore, in the setting process, different learning rates are set, and the influence of different learning rates on the recognition effect is compared, and finally the one with the smallest influence effect is selected. The smallest influence effect is based on the smallest reduction in recognition rate.
[0052] In the embodiment, the partial discharge signal acquisition platform is used to obtain the ultrasonic signal generated by the discharge. The microphone for collecting audio is placed within 1m from the outer wall of the switch cabinet, and the partial discharge phenomenon of the switch cabinet is simulated in sequence, and several discharge sounds of the switch cabinet are collected. Each type of data collected is randomly divided by audio processing software, and the waveforms generated by several discharges are analyzed at a higher resolution.
[0053] The bidirectional recurrent neural network model is trained, first, the Hilbert transform is used to envelope process the ultrasonic defect signal generated by partial discharge, analyze the differences of different types of defects, transform a continuous time domain signal into a linear output signal with impulse response, and analyze the characteristics of several defect signals, calculate the envelope signal to obtain mean, standard deviation, information entropy and other characteristic parameters, and construct a database of defect signals.
[0054] When ultrasonic waves propagate in an elastic medium, the intensity of the sound waves decreases with the increase of the propagation distance, thereby causing ultrasonic attenuation. At present, the research of engineering nondestructive testing mainly considers absorption attenuation and scattering attenuation, and does not consider diffusion attenuation, so the relationship between sound pressure p and ultrasonic propagation distance r is as follows:
[0055] ;
[0056] In the formula, r is the sound wave propagation distance (r is not in the formula); is the attenuation coefficient;
[0057] Using the above sound attenuation formula, the ultrasonic defect detection sound field theory can be converted from the sound pressure value to the signal amplitude and attenuated in a certain axis direction as follows:
[0058] ;
[0059] In the formula, is the signal amplitude in the material from the sound source excitation; alpha is the attenuation coefficient of the medium; is the signal amplitude; y is the vertical distance from the ultrasonic probe to the sound wave position.
[0060] The Hilbert transform is to transform a continuous time domain signal into a linear signal with impulse response of the signal and output the response signal. The Hilbert transform transforms the initial signal s(t) into the output signal corresponding as follows:
[0061] ;
[0062] In the formula, ;
[0063] The defect amplitudes at different times obtained by detection need to be Hilbert transformed and envelope processed as follows:
[0064] ;
[0065] In the formula, is the time domain distribution of the signal after Hilbert transform and envelope processing; is the time region of the signal; is the initial signal.
[0066] In order to effectively characterize the original defect signal, statistical feature calculation is performed on the envelope time domain signal, the mean and variance of the time domain signal can be extracted, and the specific calculation is as follows:
[0067] ;
[0068] ;
[0069] In the formula: is the positive value of the time domain amplitude of the ith sampling point; is the mean of the defect time domain signal; is the standard deviation of the defect time domain signal
[0070] The steps described above are performed before signal feature fusion, which ensures the short-time stationarity of the signal, and then the pre-processing is performed. The application of this step needs to collect field data for model training whether in training or actual discharge signal recognition. After the model training is completed, the actual discharge signal in the same place after the above steps, pre-processing and feature parameter fusion can be directly input into the recognition model without training.
[0071] In the bidirectional recurrent neural network model in this embodiment, the output value h t at time t is jointly acted by the forward hidden layer output and the reverse hidden layer output , and the calculation process can be represented as:
[0072] ;
[0073] ;
[0074] ;
[0075] In the formula, U1 and w1 are the weight matrix of the hidden layer in the forward propagation process respectively, and b1 is the bias term of the forward propagation; U2 and w2 are the weight matrix of the hidden layer in the backward propagation process respectively, and b2 is the bias term of the backward propagation.
[0076] Wherein, the Bi-GRU contains two GRU units, the forward GRU is responsible for processing from the beginning to the end of the feature sequence (i.e. from beginning to end), and the reverse GRU is responsible for processing the feature sequence in reverse (from the end to the beginning of the feature sequence). The output of the Bi-GRU layer is jointly determined by the state of the forward GRU at time t and the state of the reverse GRU at time t. The output of the Bi-GRU at each time contains the forward (past) and reverse (future) context information, and the formula used is consistent with the above formula.
[0077] The recurrent neural network can calculate the output value of the reference model using forward propagation during training, and update the parameters of the network structure according to the error using back propagation; the bidirectional recurrent neural network is formed by copying the unfolded RNN twice to form two networks in the forward and backward directions; and the BiGRU is self-updated by the updated weight values of the neural network to find the optimal weight values, so that the recognition effect can be obtained only by inputting the features. The reason for such setting is that the model has good recognition effect on signals with too long feature sequences.
[0078] The output layer of the neural network outputs three kinds of discharge signals, respectively corona discharge, surface discharge and internal discharge. In order to identify the three signals, a cross-entropy loss function suitable for multi-classification problems is used, and the recognition labels of the three discharges are set as [1, 0, 0], [0, 1, 0] and [0, 0, 1], and the loss function is:
[0079] ;
[0080] In the formula, LOSS is the loss function, N is the number of samples, Y is the expected output value, and y is the actual output value.
[0081] The following table shows the fault recognition rate obtained after inputting the MFCC, LPCC, LPCC-MFCC parameters into the recognition model, which is used to prove that the fused LPCC-MGCC parameters can improve the recognition of fault signals. The comparison of LPCC and MFCC is also based on the above J Fisher Criterion to output the features with strong representation ability, and the J Fisher Comparison of the fused LPCC-MFCC parameters.
[0082] Table 1 Recognition effect after fusion of parameters
[0083]
[0084] The above embodiments are only illustrative of the principles of the present application and its efficacy, and are not intended to limit the present application. Any modification or change made by any person skilled in the art without departing from the spirit and scope of the present application shall be covered by the claims of the present application.
Claims
1. A method for identifying partial discharge in switchgear, characterized in that, Includes the following steps: S1. Acquire the ultrasonic signal from the switch cabinet and preprocess the ultrasonic signal; S2. Convert the preprocessed ultrasound signal into a speech-like signal and extract the features of the speech-like signal; S3. Using the extracted features as input, the partial discharge type is obtained through the recognition model.
2. The method for identifying partial discharge in a switchgear according to claim 1, characterized in that, The preprocessing process includes: framing and windowing the ultrasound signal, converting the ultrasound signal into a speech-like signal using the relationship between frequency and Mel frequency, stretching subtle low-frequency components, and compressing high-frequency components.
3. The method for identifying partial discharge in a switchgear according to claim 1, characterized in that, Step S2 includes: Frequency profile, harmonics, and formants of speech-like signals are extracted using LPCC. Pitch information of speech signals is extracted using MFCC; By fusing the signal features carried by LPCC and MFCC using the Fisher criterion, the more acoustically significant LPCC-MFCC features are extracted to form a feature sequence.
4. The method for identifying partial discharge in a switchgear according to claim 2, characterized in that, The Fisher criterion is as follows: ； ； ； in, Inter-class dispersion; Within-class dispersion; Fisher's ratio; Let be the mean of the k-th dimension feature component of the i-th type signal; The number of samples for class i signals; Let be the characteristic parameters of the i-th type of signal in the k-th dimension.
5. The method for identifying partial discharge in a switchgear according to claim 1, characterized in that, The recognition model is a bidirectional recurrent neural network model.
6. The method for identifying partial discharge in a switchgear according to claim 5, characterized in that, The bidirectional recurrent neural network model includes two layers of inverse gated recurrent units as two hidden layers.