Circuit breaker mechanical fault voiceprint feature extraction method

Abstract

The invention provides a circuit breaker mechanical fault voiceprint feature extraction method, which belongs to the technical field of circuit breaker mechanical fault voiceprint, and comprises the following steps: collecting sound signals of a circuit breaker in normal operation and various mechanical fault states; performing filtering and noise reduction on the collected sound signal to obtain a preprocessed sound signal, and extracting an abnormal waveform segment set; processing the preprocessed sound signal by using a CUDA parallel processing stream of the GPU, and processing an abnormal waveform segment set by using an input stream of the CPU; performing time-frequency domain feature extraction on the preprocessed sound signal to obtain a first feature matrix; clustering the first feature matrix to obtain a plurality of clustering centers as typical sound features, and constructing a second feature matrix; extracting features of the abnormal waveform segment set to obtain a third feature matrix, and performing feature fusion and dimension reduction on the third feature matrix and the second feature matrix; and finally, a mechanical fault voiceprint feature matrix is obtained. The technical problem that a traditional serial processing mode is low in efficiency is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit breaker mechanical fault soundprint, and in particular, relates to a circuit breaker mechanical fault soundprint feature extraction method. Background Art

[0002] In modern power systems, circuit breakers are key switching devices, and their reliability directly affects the stability and safety of power supply. With the continuous expansion of the scale of power grids and the continuous growth of electricity demand, higher requirements are placed on the operating status monitoring and fault diagnosis of circuit breakers. Traditional circuit breaker fault diagnosis methods mainly rely on regular maintenance and manual inspections, which is not only time-consuming and labor-intensive, but also difficult to detect potential fault hazards in a timely manner. In recent years, fault diagnosis technology based on sound signal analysis has gradually become a research hotspot due to its advantages such as non-invasiveness, real-time and economy.

[0003] At present, the analysis of circuit breaker sound signals mainly focuses on time domain analysis, frequency domain analysis and time-frequency analysis. Time domain analysis mainly focuses on the characteristics of sound signals such as amplitude and duration, but its ability to identify complex faults is limited. Frequency domain analysis extracts spectral features through methods such as Fourier transform, but it is difficult to reflect the time-varying characteristics of the signal. Although time-frequency analysis such as short-time Fourier transform and wavelet transform can provide time domain and frequency domain information at the same time, they still have limitations when processing non-stationary signals. In addition, existing methods often only focus on a single type of feature and ignore the relationship between different features, resulting in insufficient accuracy and reliability of fault diagnosis.

[0004] In practical applications, the sound signals of circuit breakers often contain complex background noise and transient interference, which brings great challenges to feature extraction. The noise reduction methods in the existing technology often over-smooth the signal, resulting in the loss of useful fault information. On the other hand, due to the diversity and complexity of mechanical faults of circuit breakers, a single feature extraction method is difficult to fully reflect all aspects of the fault, affecting the comprehensiveness and accuracy of diagnosis.

[0005] In addition, as the amount of data increases, the traditional circuit breaker mechanical fault voiceprint feature extraction method often uses a serial method with low efficiency and can no longer meet the needs of real-time diagnosis. How to effectively utilize the parallel processing capabilities of modern computing hardware to improve the efficiency of feature extraction has also become an urgent problem to be solved.

[0006] Therefore, there is an urgent need for a method that can comprehensively utilize multiple feature information to efficiently and accurately extract features from circuit breaker sound signals under complex backgrounds, so as to improve the accuracy and real-time performance of mechanical fault diagnosis. Summary of the invention

[0007] In view of this, the present invention provides a method for extracting voiceprint features of mechanical faults of circuit breaker, which can solve the technical problem that traditional methods for extracting voiceprint features of mechanical faults of circuit breaker often use a serial method and have low efficiency.

[0008] The present invention is achieved in that:

[0009] A first aspect of the present invention provides a method for extracting voiceprint features of a circuit breaker mechanical fault, comprising the following steps:

[0010] S10, continuously collect sound signals of the circuit breaker under normal operation and various mechanical fault conditions;

[0011] S20, filtering and denoising the sound signal to obtain a preprocessed sound signal, and abnormal sudden change segments of the sound signal waveform excluded in the filtering and denoising process, recorded as an abnormal waveform segment set;

[0012] S30, starting the CUDA parallel processing flow of the GPU to receive the pre-processed sound signal; starting the input flow of the CPU to receive the abnormal waveform segment set;

[0013] S40, using a sliding window to perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis, and wavelet packet decomposition on the preprocessed sound signal to obtain features of the preprocessed sound signal, which are recorded as a first feature matrix;

[0014] S50, clustering the first feature matrix to obtain a plurality of cluster centers for representing typical sound features;

[0015] S60, for each element in the cluster, using the cluster center to replace the element to obtain a second feature matrix;

[0016] S70, extracting features of the abnormal waveform segment set to obtain a third feature matrix, and aligning the second feature matrix and the third feature matrix based on the second feature matrix;

[0017] S80, fusing the second feature matrix and the third feature matrix using a preset feature fusion equation group, and reducing the dimension using a principal component analysis method to obtain a feature matrix after dimension reduction;

[0018] S90. Standardize the feature matrix after dimensionality reduction to finally obtain a mechanical fault voiceprint feature matrix.

[0019] Specifically, the step S10 includes:

[0020] Step 101: Select a high-sensitivity microphone array as a sound collection device, and install the high-sensitivity microphone array at key positions around the circuit breaker;

[0021] Step 102, setting the sampling frequency of the high-sensitivity microphone array to 44.1 kHz, and setting the acquisition time to 5 seconds after each operation;

[0022] Step 103: Use a synchronous trigger device to ensure that the collection of the sound signal starts synchronously with the action of the circuit breaker;

[0023] Step 104, collect samples multiple times under different operating conditions, including normal operating state, slight wear state, severe wear state, and mechanical jam state, and collect samples at least 100 times in each state;

[0024] Step 105, saving the collected sound signal in WAV format;

[0025] Step 106: Establish a database to record metadata such as environmental conditions, circuit breaker status, and collection time for each collection.

[0026] Wherein, the step S20 specifically includes:

[0027] Step 201: Filter the signal using a bandpass filter, where the passband range of the filter is set to 100 Hz to 10 kHz;

[0028] Step 202: adopt Butterworth filter design, and select the order as 8th order;

[0029] Step 203, using the wavelet threshold denoising method to perform denoising, selecting the db4 wavelet basis, and setting the number of decomposition layers to 5;

[0030] Step 204: The wavelet coefficients of each layer are processed using a soft threshold function, and the threshold is selected using the Stein unbiased risk estimation criterion;

[0031] Step 205: In the noise reduction process, detect the mutation segment of the signal, define a mutation segment detection function, and when the detection function exceeds a preset threshold, mark the segment of the signal as an abnormal mutation segment;

[0032] Step 206: record the start time, duration and amplitude information of the abnormal mutation segment, exclude the abnormal mutation segment from the preprocessed signal, and save it separately as an abnormal waveform segment set.

[0033] Wherein, the step S30 specifically includes:

[0034] Step 301, initialize the NVIDIA CUDA environment, set GPU device parameters, including setting the thread block size to 256 and dynamically adjusting the grid size according to the data volume;

[0035] Step 302: Create a CUDA stream object for managing parallel tasks on the GPU;

[0036] Step 303: copy the preprocessed sound signal data from the host memory to the global memory of the GPU, and use the CUDAMemcpyAsync function to implement asynchronous data transmission;

[0037] Step 304: Create an input stream object on the CPU side to receive the abnormal waveform segment set. The CPU input stream adopts a double buffer mechanism and sets the buffer size to 1 megabyte.

[0038] Step 305: Implement a synchronization mechanism, using condition variables and mutex locks to coordinate the data processing progress of the GPU and CPU;

[0039] Step 306: Implementing a CUDA kernel function on the GPU side to parallelly process short-time Fourier transform and other operations of the preprocessed signal;

[0040] Step 307: Implement a multi-thread processing function on the CPU side to process feature extraction of abnormal waveform segment sets, and set the number of CPU threads to twice the number of system logic cores.

[0041] Wherein, the step S40 specifically includes:

[0042] Step 401, perform short-time Fourier transform, select Hanning window as window function, set the window length to 1024 sampling points, and the overlap rate to 50%;

[0043] Step 402, calculate the Mel-frequency cepstral coefficients, first convert the short-time Fourier transform result to the Mel-frequency scale, and use 40 Mel filter banks;

[0044] Step 403: Take the logarithm of the Mel filter output and perform discrete cosine transform, retaining the first 13 coefficients as Mel frequency cepstral coefficient features;

[0045] Step 404, perform wavelet packet decomposition, select db4 wavelet basis, decompose the layers to 5, and calculate the energy characteristics for each sub-band;

[0046] Step 405: combine the short-time Fourier transform, the Mel-frequency cepstral coefficients and the wavelet packet energy features into a feature matrix, which is called the first feature matrix;

[0047] Step 406: normalize the first feature matrix using a Z-score normalization method.

[0048] Wherein, the step S50 specifically includes:

[0049] Step 501: clustering is performed using an improved K-means++ algorithm, and the optimal number of clusters K is determined using the elbow rule;

[0050] Step 502, calculating the total intra-class sum of squares under different K values, and selecting the inflection point where the rate of decrease of the total intra-class sum of squares is significantly slowed down as the optimal K value;

[0051] Step 503: When initializing the cluster center, the first center is randomly selected, and the probability of selecting subsequent centers is proportional to the square of the minimum distance from the selected center;

[0052] Step 504: To improve the stability of clustering, a strategy of taking an average of multiple runs is adopted, specifically 10 runs are performed, and a result with the smallest sum of squares within the overall class is selected;

[0053] Step 505: In each iteration, the distance from each sample to its nearest center is calculated, and samples with a distance greater than 3 times the standard deviation are marked as potential outliers, and given a smaller weight when updating the cluster center;

[0054] Step 506: After clustering is completed, the representativeness score of each cluster center is calculated, and several cluster centers with the highest scores are selected as typical sound features.

[0055] Wherein, the step S60 specifically includes:

[0056] Step 601, calculating the distance from each sample point to each cluster center, using Euclidean distance for distance calculation;

[0057] Step 602: for each sample point, find the nearest cluster center, and replace the coordinates of the original sample point with the coordinates of the cluster center;

[0058] Step 603: introduce a soft allocation mechanism, define soft allocation weights, and calculate the new coordinates of the sample point as the weighted sum of the coordinates of all cluster centers;

[0059] Step 604: perform principal component analysis on the replaced feature matrix to calculate the covariance matrix;

[0060] Step 605, solve the characteristic equation, obtain the eigenvalue and eigenvector, and select the first k principal components whose cumulative contribution rate reaches 95%;

[0061] Step 606: Project the data into the subspace formed by the k principal components to obtain a second feature matrix;

[0062] Step 607: normalize the second characteristic matrix using the Min-Max normalization method.

[0063] Wherein, the step S70 specifically includes:

[0064] Step 701, extracting features from each waveform segment in the abnormal waveform segment set, using the same method as step S40 to calculate short-time Fourier transform, Mel-frequency cepstrum coefficients and wavelet packet energy features;

[0065] Step 702: make corresponding adjustments to the calculation of short-time Fourier transform and Mel-frequency cepstral coefficients, reduce the window length to 512 sampling points, and increase the overlap rate to 75%;

[0066] Step 703: for wavelet packet decomposition, keep the 5-layer decomposition, but only select the energy features of the high-frequency sub-band to highlight the abnormal features;

[0067] Step 704: Combining the extracted features into a third feature matrix;

[0068] Step 705: Use the timestamp information of the abnormal waveform segment to map it to the time frame of the second feature matrix for time alignment;

[0069] Step 706: upsample or downsample the third feature matrix using a linear interpolation method to make its dimension consistent with the second feature matrix, and perform feature dimension alignment;

[0070] Step 707: normalize the third characteristic matrix using the same Min-Max normalization method as the second characteristic matrix.

[0071] Wherein, the step S80 specifically includes:

[0072] Step 801: Apply a preset feature fusion equation group to perform feature fusion, including time-frequency domain fusion, energy distribution fusion, phase spectrum fusion and abnormal feature fusion;

[0073] Step 802: Use the particle swarm optimization algorithm to optimize the weight coefficient in the fusion equation, with the population size set to 50 and the maximum number of iterations set to 100;

[0074] Step 803: The optimization objective function is the ratio of the inter-class distance to the intra-class distance of the fused features;

[0075] Step 804: performing centralization processing on the fused feature matrix;

[0076] Step 805, calculate the covariance matrix, solve the characteristic equation, and obtain the eigenvalue and the corresponding eigenvector;

[0077] Step 806: Arrange the eigenvalues ​​in descending order and select the first k principal components whose cumulative contribution rate reaches 95%;

[0078] Step 807: construct a projection matrix, whose columns are selected eigenvectors, and obtain a reduced-dimensional eigenmatrix through matrix multiplication.

[0079] Wherein, the step S90 specifically includes:

[0080] Step 901: Use the box plot method to identify outliers, and consider values ​​that are less than the first quartile minus 1.5 times the interquartile range or greater than the third quartile plus 1.5 times the interquartile range as potential outliers;

[0081] Step 902: Use the tail-shrinking method to limit the detected outliers to the range of the first quartile minus 1.5 times the interquartile range to the third quartile plus 1.5 times the interquartile range;

[0082] Step 903, performing Z-score standardization on the processed feature matrix;

[0083] Step 904: Calculate the correlation matrix using the Pearson correlation coefficient to check the correlation between the features;

[0084] Step 905: If there is a highly correlated feature pair, and the absolute value of the correlation coefficient is greater than 0.9, consider deleting one of the features to reduce the multicollinearity problem;

[0085] Step 906: Name and describe the standardized features, and name them according to their sources and physical meanings;

[0086] Step 907: construct a feature importance ranking, based on the variance contribution rate of the feature or the feature importance score using a model such as a random forest, and append the feature importance information to the final feature matrix.

[0087] The clustering is implemented using an improved K-means++ algorithm.

[0088] Further, 1. Time-frequency domain fusion equation:

[0089]

[0090] Among them: F t (x, y) is the fused time-frequency domain feature; S(x, y) is the time-frequency domain feature in the second feature matrix; T(x, y) is the time-frequency domain feature in the third feature matrix; α, β, γ, δ are weight coefficients, which are obtained through the particle swarm optimization algorithm; is the second-order partial derivative of S(x,y), which is used to enhance edge information; A(x,y) is the time-frequency feature of the abnormal mutation segment, which is obtained by short-time Fourier transform.

[0091] 2. Energy distribution fusion equation:

[0092]

[0093] Where: E(f) is the energy distribution characteristic after fusion; P i (f) is the power spectral density of the i-th feature matrix; w i is the weight coefficient, calculated by entropy weight method; N is the number of feature matrices (here 2); η is the weight coefficient of abnormal mutation segment energy; M is the number of abnormal mutation segments; D j is the duration of the jth abnormal mutation segment; T j is the total duration of the signal; E j (f) is the energy spectrum density of the jth abnormal mutation segment.

[0094] 3. Phase spectrum fusion equation:

[0095]

[0096] Where: Φ(f) is the phase spectrum feature after fusion; φ i (f) is the phase spectrum of the i-th characteristic matrix; v i is the weight coefficient, obtained through phase consistency evaluation; k is an integer used to ensure phase continuity; μ is the weight coefficient of the abnormal mutation phase; A j is the mutation amplitude of the jth abnormal mutation segment; A max is the maximum mutation amplitude among all abnormal mutation segments; φ j (f) is the phase spectrum of the jth abnormal mutation segment.

[0097] 4. Abnormal feature fusion equation:

[0098]

[0099] Among them: F a (t) is the abnormal feature after fusion; M is the number of abnormal mutation segments; λ j is the weight coefficient of the jth abnormal mutation segment, which is calculated comprehensively by the mutation amplitude and duration; t j,start and t j,end are the start and end time of the jth abnormal mutation segment respectively; t j,mid is the midpoint of the jth abnormal mutation segment; σ j is the time scale parameter of the jth abnormal mutation segment, which is related to the duration; A j is the mutation amplitude of the jth abnormal mutation segment.

[0100] in,

[0101] Among them, s a (t) is the abnormal mutation segment signal, w(t) is the window function (such as Hanning window).

[0102]

[0103] Among them, s i (t) is the time domain signal corresponding to the i-th feature matrix, and T is the signal duration.

[0104]

[0105] Where, Hi = -∑fpi(f)logpi(f),

[0106]

[0107] Among them, s j (t) is the signal of the jth abnormal mutation segment.

[0108]

[0109] in, is the average phase of all eigenmatrices at frequency f.

[0110]

[0111] Among them, X i (f) is the Fourier transform of the i-th feature matrix.

[0112]

[0113]

[0114] This setting makes the value of the Gaussian function at the beginning and end of the abnormal mutation segment approximately 1% of the peak value.

[0115]

[0116] D=tj,endd-tj,start

[0117]

[0118] in, is the local mean or trend of the signal.

[0119]

[0120] These calculation formulas provide a method to obtain each variable. In practical applications, appropriate adjustments and optimizations may be required based on specific signal characteristics and system requirements. For example, the weight coefficients (α, β, γ, δ, η, μ) can be optimized by machine learning algorithms (such as genetic algorithms or particle swarm optimization) to obtain the best fusion effect.

[0121] Compared with the prior art, the method for extracting the voiceprint feature of a circuit breaker mechanical fault provided by the present invention has the following beneficial effects:

[0122] First, the present invention adopts a multimodal feature fusion method, comprehensively utilizing time-frequency domain features, energy distribution features, phase spectrum features, and abnormal features to fully capture all aspects of the circuit breaker sound signal. This multi-dimensional feature extraction method greatly enhances the ability to identify complex mechanical faults and overcomes the limitations of a single feature extraction method.

[0123] Secondly, the present invention introduces innovative preprocessing technology. By combining bandpass filtering and wavelet threshold noise reduction, the background noise is effectively removed while retaining key fault information. In particular, the separate processing of abnormal mutation segments ensures that transient fault information will not be lost in the noise reduction process, greatly improving the accuracy of feature extraction.

[0124] Thirdly, the present invention adopts a GPU-CPU collaborative processing architecture, which fully utilizes the parallel computing capabilities of modern hardware. The GPU is responsible for parallel processing of large-scale data, such as short-time Fourier transform and other computationally intensive tasks, while the CPU handles tasks that require flexible control, such as abnormal waveform segments. This collaborative processing mechanism significantly improves the efficiency of feature extraction, making it possible to monitor large-scale circuit breaker groups in real time.

[0125] In addition, the present invention introduces cluster analysis and soft allocation mechanism, which effectively reduces the redundancy of data and extracts the most representative sound features. This not only reduces the amount of calculation for subsequent processing, but also improves the expressiveness of features. At the same time, by reducing the dimension through principal component analysis, the feature space is further optimized, which improves the efficiency and accuracy of subsequent fault diagnosis.

[0126] Finally, the present invention adopts an innovative feature fusion equation group, including time-frequency domain fusion, energy distribution fusion, phase spectrum fusion and abnormal feature fusion. This multi-level fusion strategy fully utilizes the advantages of various features and significantly improves the ability to distinguish features. In particular, the introduction of abnormal feature fusion equations effectively enhances the ability to identify transient faults.

[0127] In general, the method of the present invention solves the technical problem that the traditional circuit breaker mechanical fault voiceprint feature extraction method often uses a serial method and has low efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0128] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0129] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0130] like Figure 1 As shown, it is a flow chart of a method for extracting voiceprint features of a circuit breaker mechanical fault provided by the present invention. The method comprises the following steps:

[0131] S10, continuously collect sound signals of the circuit breaker under normal operation and various mechanical fault conditions;

[0132] S20, filtering and denoising the sound signal to obtain a preprocessed sound signal and abnormal sudden change segments of the sound signal waveform excluded during the filtering and denoising process, recorded as an abnormal waveform segment set;

[0133] S30, starting the CUDA parallel processing flow of the GPU to receive the preprocessed sound signal; starting the input flow of the CPU to receive the abnormal waveform segment set;

[0134] S40, using a sliding window to perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis, and wavelet packet decomposition on the preprocessed sound signal to obtain the characteristics of the preprocessed sound signal, which are recorded as the first characteristic matrix;

[0135] S50, clustering the first feature matrix to obtain a plurality of cluster centers for representing typical sound features;

[0136] S60, for each element in the cluster, using the cluster center to replace the element, to obtain a second feature matrix;

[0137] S70, extracting features of the abnormal waveform segment set to obtain a third feature matrix, and aligning the second feature matrix and the third feature matrix based on the second feature matrix;

[0138] S80, fusing the second feature matrix and the third feature matrix using a preset feature fusion equation group, and reducing the dimension using a principal component analysis method to obtain a feature matrix after dimension reduction;

[0139] S90, standardize the feature matrix after dimensionality reduction, and finally obtain the mechanical fault voiceprint feature matrix.

[0140] The specific implementation methods of the above steps are described in detail below:

[0141] The specific implementation method of step S10 is: first, select a suitable sound collection device, such as a high-sensitivity microphone array, and install it at key positions around the circuit breaker. The sampling frequency of the collection device is set to 44.1kHz to ensure that all sound signals within the audible range of the human ear can be captured. The collection time is set to 5 seconds after each operation, which is sufficient to cover the complete action process of the circuit breaker. During the collection process, a synchronous trigger device is used to ensure that the collection of sound signals starts synchronously with the action of the circuit breaker. In order to obtain a comprehensive data set, multiple collections are required under different operating conditions, including normal operating conditions, slight wear conditions, severe wear conditions, mechanical jamming conditions, etc. At least 100 samples are collected for each state to ensure the statistical significance of the data. The collected sound signals are saved in WAV format to ensure the integrity of the original data. At the same time, a detailed database is established to record metadata such as environmental conditions, circuit breaker status, and collection time for each collection, providing important reference information for subsequent data analysis. The main purpose of this step is to obtain comprehensive and accurate original sound data, laying the foundation for subsequent feature extraction and fault diagnosis.

[0142] The specific implementation method of step S20 is: pre-process the collected original sound signal to improve the quality and reliability of the signal. First, use a bandpass filter to filter the signal, and the passband range of the filter is set to 100Hz to 10kHz. This range covers the frequency components of most mechanical fault sounds. The filter is designed using a Butterworth filter, and the order is selected to be 8th order to ensure good roll-off characteristics. After filtering, the wavelet threshold denoising method is used for noise reduction. The db4 wavelet basis is selected, and the number of decomposition layers is 5. The wavelet coefficients of each layer are processed using a soft threshold function, and the threshold is selected using the Stein unbiased risk estimation criterion. The specific formula is: Where σ is the standard deviation of noise and N is the length of the signal. In the process of noise reduction, the mutation segment of the signal is detected at the same time. Define the mutation segment detection function Where s(t) is the original signal, is the local mean of the signal. When D(t) exceeds the preset threshold (recommended to be set to 3 times the standard deviation of the signal), the signal segment is marked as an abnormal mutation segment, and its start time, duration, and amplitude information are recorded. These abnormal mutation segments are excluded from the preprocessed signal and saved separately as an abnormal waveform segment set for subsequent feature fusion. The preprocessed signal and the abnormal waveform segment set are saved as independent data files for subsequent processing. The main purpose of this step is to improve the signal-to-noise ratio of the signal, remove irrelevant noise interference, and retain abnormal mutation features that may contain important fault information.

[0143] The specific implementation method of step S30 is: in order to improve the efficiency of data processing, this step adopts the collaborative working mode of GPU and CPU. First, initialize the NVIDIA CUDA environment, set the GPU device parameters, including the thread block size (recommended to be set to 256) and the grid size (dynamically adjusted according to the amount of data). Create a CUDA stream object for managing parallel tasks on the GPU. The preprocessed sound signal data is copied from the host memory to the global memory of the GPU, and the CUDAMemcpyAsync function is used to implement asynchronous data transmission to maximize the GPU utilization. At the same time, an input stream object is created on the CPU side for receiving an abnormal waveform segment set. The CPU input stream adopts a double buffering mechanism, and the buffer size is set to 1MB to balance memory usage and data reading efficiency. In order to ensure the integrity and consistency of the data, a synchronization mechanism is implemented, and conditional variables and mutexes are used to coordinate the data processing progress of the GPU and CPU. On the GPU side, a CUDA kernel function is implemented for parallel processing of short-time Fourier transform and other operations of preprocessed signals, preferably the Fourier transform function of CUDA. On the CPU side, a multi-threaded processing function is implemented to process the feature extraction of abnormal waveform segment sets, and a multi-threaded Fourier transform function is preferably used. The number of CPU threads is set to twice the number of system logic cores to make full use of CPU resources. Through this GPU-CPU collaborative processing method, the speed of data processing can be significantly improved, especially for large-scale data sets. Especially in actual use, the number of abnormal mutation segments filtered is small but the spectral characteristics are relatively complex, and the number of pre-processed sound signals after filtering is large but the spectral characteristics are relatively simple. Therefore, using the CPU to process a small amount of abnormal mutation segments with relatively complex spectral characteristics, and using the GPU to process a large number of pre-processed sound signals with relatively simple spectral characteristics can well distribute GPU-CPU collaborative processing and improve overall processing efficiency. The main purpose of this step is to prepare an efficient parallel computing environment for the subsequent feature extraction process and improve the processing efficiency of the entire system.

[0144] The specific implementation of step S40 is: using the parallel processing capability of the GPU to extract multiple features from the pre-processed sound signal. First, a short-time Fourier transform (STFT) is performed. The Hanning window is selected as the window function, the window length is set to 1024 sampling points, and the overlap rate is 50%. The calculation formula of STFT is: Where x[n] is the input signal, w[n] is the window function, m is the time index, and ω is the angular frequency. Next, the Mel-frequency cepstral coefficients (MFCC) are calculated. First, the STFT result is converted to the Mel frequency scale using a 40-Mel filter bank. The Mel frequency conversion formula is: Then, the logarithm of the filter output is taken and discrete cosine transform is performed, and the first 13 coefficients are retained as MFCC features. Finally, wavelet packet decomposition is performed. The db4 wavelet basis is selected and the number of decomposition layers is 5. The energy feature E is calculated for each subband. i =∑ k |d i [k]| 2 , where d i [k] is the wavelet coefficient of the ith subband. The STFT, MFCC and wavelet packet energy features are combined into a feature matrix, called the first feature matrix. Each row of the matrix represents a time frame, and the columns contain different types of features. In order to improve the distinguishability of the features, the first feature matrix is ​​normalized using the Z-score normalization method: Where μ is the feature mean and σ is the feature standard deviation. The main purpose of this step is to extract multi-dimensional and multi-scale sound features from the preprocessed signal to provide rich feature representation for subsequent fault identification.

[0145] The specific implementation method of step S50 is: cluster analysis is performed on the first feature matrix to extract typical sound features. The improved K-means++ algorithm is used for clustering. First, the optimal number of clusters K is determined using the elbow rule. The overall intra-class sum of squares (WSS) under different K values ​​is calculated, and the inflection point where the WSS decline rate is significantly slowed down is selected as the optimal K value, which is generally between 5 and 10. When initializing the cluster center, the first center is randomly selected, and the selection probability of the subsequent center is proportional to the square of the minimum distance from the selected center. The iterative update formula of the clustering process is: where c i is the i-th cluster center, S i is the set of samples belonging to the i-th class. In order to improve the stability of clustering, the strategy of taking the average of multiple runs is adopted. Specifically, it is run 10 times and the result with the smallest sum of squares within the overall class is selected. In order to deal with possible outliers, in each iteration, the distance from each sample to its nearest center is calculated, and samples with a distance greater than 3 times the standard deviation are marked as potential outliers, and a smaller weight is given when updating the cluster center. After clustering is completed, the representative score of each cluster center is calculated. The scoring formula is: where σ i is the average distance of samples in the i-th class. Select several cluster centers with the highest scores as typical sound features. The main purpose of this step is to extract the most representative sound feature pattern from a large amount of feature data, simplify the subsequent feature matching process, and improve the efficiency and accuracy of fault identification.

[0146] The specific implementation of step S60 is: based on the cluster centers obtained in step S50, the first feature matrix is ​​subjected to dimensionality reduction and feature replacement. First, the distance from each sample point to each cluster center is calculated, and the distance calculation adopts the Euclidean distance: Where x is the sample point, c i is the i-th cluster center, and n is the feature dimension. Then, for each sample point, find the nearest cluster center and replace the coordinates of the original sample point with the coordinates of the cluster center. In order to retain certain local structural information, a soft allocation mechanism is introduced. Define the soft allocation weight: Where T is the temperature parameter, which controls the degree of soft assignment and is recommended to be set to the average value of the distance. The new coordinates of the sample point are calculated as: This soft allocation mechanism can retain the main features while also partially retaining the information of the secondary features. In order to further reduce the redundancy of the data, the principal component analysis (PCA) is performed on the replaced feature matrix. Calculate the covariance matrix: Where m is the number of samples, is the sample mean. Solve the characteristic equation Cv=γv to obtain the eigenvalue and eigenvector. Select the first k principal components with a cumulative contribution rate of 95%, project the data into the subspace formed by these k principal components, and obtain the second characteristic matrix. Finally, normalize the second characteristic matrix using the Min-Max normalization method: The main purpose of this step is to significantly reduce the amount of data and improve the efficiency of subsequent processing while retaining key information through feature replacement and dimensionality reduction.

[0147] The specific implementation method of step S70 is: process the abnormal waveform segment set and align its features with the second feature matrix. First, feature extraction is performed on each waveform segment in the abnormal waveform segment set. The STFT, MFCC and wavelet packet energy features are calculated using the same method as step S40. Considering the short-term characteristics of the abnormal waveform segment, the calculation of STFT and MFCC is adjusted accordingly: the window length is reduced to 512 sampling points, and the overlap rate is increased to 75%. For wavelet packet decomposition, the 5-layer decomposition is maintained, but only the energy features of the high-frequency subband (the 4th and 5th layers) are selected to highlight the abnormal features. The extracted features are combined into a third feature matrix. In order to align with the second feature matrix, alignment of time and feature dimensions is required. Time alignment: Use the timestamp information of the abnormal waveform segment to map it to the time frame of the second feature matrix. If an abnormal waveform segment spans multiple time frames, its features are distributed to the corresponding frames according to the time ratio. Feature dimension alignment: Since the third feature matrix may have a different dimension from the second feature matrix, dimensional adjustment is required. The third feature matrix is ​​upsampled or downsampled using the linear interpolation method to make its dimension consistent with the second feature matrix. The interpolation formula is: Among them, (x1, y1) and (x2, y2) are adjacent known points, (x i ,y i ) is the point to be interpolated. After alignment, the third feature matrix is ​​normalized using the same Min-Max normalization method as the second feature matrix. In order to enhance the expression of abnormal features, the attention mechanism is introduced. Calculate the abnormality score: where x i is the i-th eigenvector, is the average feature vector, and σ is the standard deviation. Construct attention weights based on the scores: The attention weights are applied to the third feature matrix to obtain the weighted feature matrix. The main purpose of this step is to effectively integrate the feature information of the abnormal waveform segment with the features of the normal signal, providing a more comprehensive information basis for subsequent feature fusion.

[0148] The specific implementation of step S80 is: fuse the second feature matrix and the third feature matrix, and perform dimensionality reduction processing. First, apply the preset feature fusion equation group to perform feature fusion. The time-frequency domain fusion adopts the equation: Among them, the initial values ​​of α, β, γ, and δ are set to 0.4, 0.3, 0.2, and 0.1, and then optimized by the particle swarm optimization algorithm. The parameter settings of particle swarm optimization are: the population size is 50, the maximum number of iterations is 100, the inertia weight w decreases linearly from 0.9 to 0.4, and the acceleration constant c1=c2=2. The optimization objective function is the ratio of the inter-class distance to the intra-class distance of the fused feature. The energy distribution fusion adopts the equation: Among them, w i The entropy weight method is used to calculate, and η is initially set to 0.2, and then optimized by cross-validation. The phase spectrum fusion uses the equation: Among them, v i Obtained through phase consistency evaluation, μ is initially set to 0.1 and optimized through grid search. The abnormal feature fusion adopts the equation: After fusion, a high-dimensional feature matrix is ​​obtained. Next, principal component analysis (PCA) is used for dimensionality reduction. First, the fused feature matrix is ​​centralized: in is the mean of each feature. Calculate the covariance matrix: Where n is the number of samples. Solve the characteristic equation Cv = λv and get the characteristic value λ i and the corresponding eigenvector v i . Arrange the eigenvalues ​​in descending order and select the first k principal components whose cumulative contribution rate reaches 95%. Contribution rate calculation formula: Construct the projection matrix P, whose columns are the selected eigenvectors. Finally, through X reduced =X centered P obtains the feature matrix after dimensionality reduction. In order to improve the robustness of PCA, the incremental PCA algorithm is considered to process large-scale data sets in batches. The update formula of incremental PCA is: Where t is the current sample number, x t The main purpose of this step is to make full use of information from different sources through feature fusion, remove redundant information through dimensionality reduction, and extract the most representative feature combination.

[0149] The specific implementation method of step S90 is: standardize the feature matrix after dimensionality reduction to obtain the final mechanical fault voiceprint feature matrix. First, check whether there are outliers in the feature matrix after dimensionality reduction. Use the box plot method to identify outliers, and regard values ​​less than Q1-1.5IQR or greater than Q3+1.5IQR as potential outliers, where Q1 and Q3 are the first and third quartiles, respectively, and IQR is the interquartile range. For the detected outliers, instead of simply deleting them, the shrinkage processing method is used to limit the outliers to the range of [Q1-1.5IQR, Q3+1.5IQR]. Then, the processed feature matrix is ​​Z-score standardized: Where μ is the feature mean and σ is the feature standard deviation. To improve the robustness of standardization, consider using the median and median absolute deviation (MAD) instead of the mean and standard deviation: Where MAD(x) = median(|x-median(x)|). After standardization, check the correlation between features and calculate the correlation matrix using the Pearson correlation coefficient. If there are highly correlated feature pairs (the absolute value of the correlation coefficient is greater than 0.9), consider deleting one of the features to reduce the problem of multicollinearity. In order to enhance the interpretability of the features, name and describe the standardized features. For example, the features can be named according to their source (time-frequency domain, energy distribution, phase spectrum, abnormal features) and physical meaning (frequency center, energy concentration, phase consistency, etc.). Finally, construct a feature importance ranking, which can be based on the variance contribution rate of the feature or the feature importance score of a model such as a random forest. The feature importance information is attached to the final feature matrix so that the subsequent fault diagnosis model can use these features more specifically. The main purpose of this step is to standardize the features after dimensionality reduction so that features of different dimensions can be directly compared. At the same time, through outlier processing and feature selection, the quality and reliability of the features can be further improved to provide high-quality input data for subsequent fault diagnosis and classification tasks.

[0150] Through the steps S10 to S90 described in detail above, the whole process from the acquisition of the original sound signal to the final mechanical fault soundprint feature matrix is ​​completed. This process covers key links such as signal acquisition, preprocessing, feature extraction, clustering analysis, feature fusion, dimensionality reduction and standardization. Each step is carefully designed to extract and retain the sound feature information related to the mechanical fault of the circuit breaker to the greatest extent. By using a variety of advanced signal processing and machine learning techniques, such as wavelet analysis, deep learning, clustering algorithm, principal component analysis, etc., the most representative and distinguishing features can be extracted from complex sound signals. The final mechanical fault soundprint feature matrix not only contains rich time-frequency domain information, energy distribution characteristics, phase characteristics and abnormal characteristics, but also undergoes dimensionality reduction and standardization processing, making it more suitable for subsequent fault diagnosis and classification tasks. The innovation of this method is that it comprehensively considers the sound characteristics under normal operation and various fault conditions, and effectively integrates this information through feature fusion. By introducing the mechanism of GPU parallel processing and CPU collaborative work, this method also has high computational efficiency and can process large-scale sound data. This systematic and comprehensive feature extraction method provides strong technical support for mechanical fault diagnosis of circuit breakers, and is expected to significantly improve the accuracy and timeliness of fault detection, making an important contribution to the safe and stable operation of power systems.

[0151] Specifically, the principle of the present invention is:

[0152] 1. Principle of multimodal feature fusion: Mechanical failures of circuit breakers often manifest as complex sound patterns, and a single type of feature is difficult to fully capture fault information. The present invention constructs a multidimensional feature space by fusing time-frequency domain features, energy distribution features, phase spectrum features, and abnormal features. The time-frequency domain features reflect the time-varying frequency characteristics of the signal, the energy distribution features characterize the energy proportion of different frequency bands, the phase spectrum features capture the phase change law of the signal, and the abnormal features focus on describing transient fault information. This multimodal fusion strategy ensures the comprehensiveness of feature extraction and can more accurately describe various complex fault modes.

[0153] 2. Adaptive signal preprocessing principle: Considering that the circuit breaker sound signal often contains complex background noise, the present invention adopts an adaptive preprocessing method. The bandpass filter is used to preliminarily remove irrelevant frequency bands, while the wavelet threshold noise reduction can adaptively adjust the threshold according to the local characteristics of the signal, effectively removing noise while retaining signal details. In particular, the separate processing of abnormal mutation segments reflects the special attention to transient fault information. This differentiated processing strategy ensures that key fault information will not be lost during the preprocessing process.

[0154] 3. Parallel computing acceleration principle: Based on the characteristics of modern computing hardware, the present invention adopts a GPU-CPU collaborative processing architecture. The large-scale parallel processing capability of the GPU is particularly suitable for computationally intensive tasks such as Fourier transform and wavelet decomposition, while the CPU is more suitable for processing tasks that require complex control processes, such as the processing of abnormal waveform segments. This collaborative processing mechanism fully utilizes the advantages of heterogeneous computing platforms and significantly improves the efficiency of feature extraction.

[0155] 4. Feature space optimization principle: In order to improve the expressiveness of features and reduce redundancy, the present invention introduces cluster analysis and principal component analysis. Cluster analysis extracts the most representative sound features by identifying similar feature patterns, thereby reducing data redundancy. The introduction of the soft allocation mechanism retains the information of the secondary features while retaining the main features, thereby improving the robustness of the features. Principal component analysis further optimizes the feature space by finding the direction of maximum variance, removes redundant information, and improves the ability to distinguish features.

[0156] 5. Principle of nonlinear feature fusion: The feature fusion equation group proposed in the present invention adopts a nonlinear fusion strategy. The time-frequency domain fusion equation introduces the second-order partial derivative to enhance the capture of edge information. The energy distribution fusion equation adopts the form of entropy to better describe the degree of energy dispersion. The phase spectrum fusion equation takes into account the cyclic characteristics of the phase and avoids the problems caused by phase jumps. The abnormal feature fusion equation adopts the form of Gaussian function, which can more accurately describe the time distribution of mutation features. This nonlinear fusion strategy can better capture the complex relationship between features and improve the expressiveness of features.

[0157] 6. Adaptive parameter optimization principle: The present invention introduces parameter optimization mechanisms in multiple links. For example, in the feature fusion process, the particle swarm optimization algorithm is used to dynamically adjust the fusion weight to adapt to different types of fault features. In cluster analysis, the elbow rule is used to adaptively determine the optimal number of clusters. This adaptive parameter optimization strategy makes the method more versatile and adaptable, and can cope with circuit breaker failures under different working conditions.

[0158] In summary, the technical solution of the present invention constructs a comprehensive, efficient and accurate circuit breaker mechanical fault voiceprint feature extraction framework through the organic combination of multimodal feature fusion, adaptive signal preprocessing, parallel computing acceleration, feature space optimization, nonlinear feature fusion and adaptive parameter optimization.

[0159] An example of a specific application scenario of the present invention is provided below: A provincial power company implemented the circuit breaker mechanical fault voiceprint feature extraction system based on the present invention in its main substation. The substation has a total of 50 10kV circuit breakers, model ZN63A-12, with service life ranging from 2 to 15 years. In order to verify the effectiveness of the method of the present invention, the power company selected 10 circuit breakers of different service life for a 6-month trial operation.

[0160] 1. Data collection stage

[0161] First, the staff installed four high-sensitivity microphones (model: PCB378B02) around each selected circuit breaker, with the sampling frequency set to 44.1kHz. The microphones were arranged in the front, back, left and right directions of the circuit breaker, about 0.5 meters away from the circuit breaker body. At the same time, a synchronous trigger system was installed to ensure that the acquisition of sound signals started synchronously with the operation of the circuit breaker.

[0162] During the six-month trial operation, the system recorded the normal operating sound of each circuit breaker and various possible fault sounds. The specific collection is as follows:

[0163] -Normal operation sound: Each circuit breaker is routinely operated at least 5 times a week, and a total of about 1,300 normal operation sounds are collected.

[0164] - Mild wear state: The contact gap is manually adjusted to simulate a mild wear state, and each circuit breaker is collected 50 times.

[0165] - Severe wear state: Severe wear state is simulated by replacing some worn parts, and each circuit breaker is collected 30 times.

[0166] -Mechanical stuck state: The mechanical stuck state is simulated by artificially increasing the friction of the operating mechanism, and each circuit breaker is collected 20 times.

[0167] All collected sound signals are saved in WAV format, and detailed metadata are recorded, including circuit breaker number, operation time, ambient temperature, humidity and other information.

[0168] 2. Data preprocessing stage

[0169] The collected original sound signal is preprocessed. First, a bandpass filter is used for filtering, and the passband range is set to 100Hz to 10kHz. The filter is designed using an 8th-order Butterworth filter. Then, a 5-layer wavelet threshold denoising method based on the db4 wavelet is used for denoising.

[0170] During the noise reduction process, the system detected multiple abnormal mutation segments. Taking one operation of a circuit breaker (No.: CB-001) as an example, three significant abnormal mutation segments were detected:

[0171] 1. Time: 0.15s-0.18s, Amplitude: 1.2 times standard deviation

[0172] 2. Time: 0.42s-0.46s, Amplitude: 2.8 times standard deviation

[0173] 3. Time: 0.89s-0.91s, Amplitude: 1.5 times standard deviation

[0174] These abnormal mutation segments are saved separately as abnormal waveform segment sets for subsequent feature fusion.

[0175] 3. Parallel processing configuration

[0176] The system is configured with a high-performance server equipped with NVIDIA Tesla V100 GPU and dual-core Intel XeonGold 6248R CPU. When the CUDA environment is initialized, the thread block size is set to 256, and the grid size is dynamically adjusted according to the amount of data. 48 threads (corresponding to 24 physical cores) are created on the CPU side to process the abnormal waveform segment set.

[0177] 4. Feature extraction stage

[0178] Perform multi-dimensional feature extraction on the preprocessed signal. Taking a normal operation of the CB-001 circuit breaker as an example, the extracted features are as follows:

[0179] Short-time Fourier transform (STFT):

[0180] -Window length: 1024 samples

[0181] - Overlap rate: 50%

[0182] -Frequency resolution: about 43Hz

[0183] -Time resolution: about 11.6ms

[0184] Mel-frequency cepstral coefficients (MFCC):

[0185] - Number of Mel filters: 40

[0186] -Retention coefficient: 13

[0187] Wavelet packet decomposition:

[0188] - Wavelet basis: db4

[0189] - Decomposition layers: 5

[0190] -Number of extracted sub-band energy features: 32

[0191] 5. Cluster analysis

[0192] Perform cluster analysis on the extracted features. Use the improved K-means++ algorithm to determine the optimal number of clusters through the elbow rule. After multiple experiments, the optimal number of clusters K=7 is determined. The clustering results are shown in Table 1:

[0193] Table 1 Clustering results

[0194] Cluster number Sample size Main features description 1 3245 Low-frequency energy is concentrated and MFCC coefficients are stable 2 2876 The intermediate frequency energy is prominent and the phase spectrum fluctuates greatly 3 1987 Rich high-frequency components and uniform wavelet energy distribution 4 1532 The time-frequency characteristics change dramatically, which may be related to mechanical jamming 5 1245 The MFCC coefficient changes significantly, which may indicate a contact problem. 6 987 The low-frequency energy decays slowly, and there may be problems with the operating mechanism. 7 654 High frequency bursts are strong and may be related to arc faults

[0195] 6. Feature replacement and dimensionality reduction

[0196] Based on the clustering results, feature replacement is performed on each sample point. A soft assignment mechanism is used, and the temperature parameter T is set to the average value of the distance (about 0.35). After feature replacement, PCA dimension reduction is performed on the data. The principal component with a cumulative contribution rate of 95% is selected, and the feature dimension is finally reduced from the original 85 dimensions to 23 dimensions.

[0197] 7. Feature Fusion

[0198] The preset feature fusion equations are used for feature fusion. Taking time-frequency domain fusion as an example, the fusion weight coefficients are obtained by particle swarm optimization algorithm: α = 0.42, β = 0.28, γ = 0.18, δ = 0.12.

[0199] In energy distribution fusion, the weight coefficient η of the abnormal mutation segment energy is set to 0.2. In phase spectrum fusion, the weight coefficient μ of the abnormal mutation segment phase is set to 0.15.

[0200] 8. Final feature matrix construction

[0201] After standardization and correlation analysis, an 18-dimensional mechanical fault voiceprint feature matrix is ​​finally obtained. The features of each dimension and their importance scores are shown in Table 2:

[0202] Table 2 Dimensional feature table

[0203]

[0204]

[0205] 9. System performance evaluation

[0206] In order to evaluate the performance of the proposed method, the differences in fault detection accuracy and calculation time between the traditional method (using only time domain and frequency domain features) and the proposed method are compared. The test data set contains 2000 normal samples and 500 various fault samples. The results are shown in Table 3:

[0207] Table 3 Performance comparison table

[0208] Performance Indicators Traditional methods Method of the present invention Improvement Fault detection accuracy 87.5％ 96.8％ 10.6％ Average computation time (per sample) 2.3s 0.8s 65.2％ False Positive Rate 8.2％ 2.5％ 69.5％ False negative rate 6.8％ 1.8％ 73.5％

[0209] 10. Typical failure case analysis

[0210] During the trial run, the system successfully detected several potential faults. The following is a typical case:

[0211] Circuit breaker number: CB-007 (service life: 8 years)

[0212] Detection time: July 15, 2024 14:32:45

[0213] Environmental conditions: temperature 28°C, humidity 65%

[0214] The system detected that the sound characteristics of the circuit breaker were abnormal. The main abnormalities were manifested in the following characteristic dimensions:

[0215] 1. The mid-frequency energy ratio (feature 2) increased significantly, from the normal level of 0.28 to 0.41.

[0216] 2. The fluctuation of the second coefficient of MFCC (feature 6) increases, and the standard deviation increases from 0.03 to 0.09.

[0217] 3. The energy of the high frequency subband of the wavelet packet (feature 9) suddenly increases, with an amplitude exceeding 2.5 times the normal level.

[0218] 4. The intensity (feature 13) and duration (feature 14) of abnormal mutations increased significantly.

[0219] Based on these abnormal characteristics, the system inferred that the circuit breaker may have excessive contact wear. The power company immediately arranged for an inspection and confirmed that the contacts were indeed severely worn. The relevant parts were replaced in time to avoid a more serious failure that might have occurred.

[0220] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for extracting voiceprint features of mechanical faults of circuit breakers, characterized in that: The following steps are involved: S10, continuously collect sound signals of the circuit breaker under normal operation and various mechanical fault conditions; S20, filtering and denoising the sound signal to obtain a preprocessed sound signal, and abnormal sudden change segments of the sound signal waveform excluded in the filtering and denoising process, recorded as an abnormal waveform segment set; S30, starting the CUDA parallel processing flow of the GPU to receive the pre-processed sound signal; Starting the input stream of the CPU to receive the abnormal waveform segment set; S40, using a sliding window to perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis, and wavelet packet decomposition on the preprocessed sound signal to obtain features of the preprocessed sound signal, which are recorded as a first feature matrix; S50, clustering the first feature matrix to obtain a plurality of cluster centers for representing typical sound features; S60, for each element in the cluster, using the cluster center to replace the element to obtain a second feature matrix; S70, extracting the features of the abnormal waveform segment set from the second feature matrix to obtain a third feature matrix, and aligning the second feature matrix and the third feature matrix based on the second feature matrix; S80, fusing the second feature matrix and the third feature matrix using a preset feature fusion equation group, and reducing the dimension using a principal component analysis method to obtain a feature matrix after dimension reduction; S90, standardize the feature matrix after dimensionality reduction, and finally obtain the mechanical fault voiceprint feature matrix.

2. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 1, characterized in that: The time-frequency domain fusion equation is specifically expressed as: Among them: F t (x, y) is the fused time-frequency domain feature; S(x, y) is the time-frequency domain feature in the second feature matrix; T(x, y) is the time-frequency domain feature in the third feature matrix; α, β, γ, δ are weight coefficients, which are obtained through the particle swarm optimization algorithm; is the second-order partial derivative of S(x,y), which is used to enhance edge information; A(x,y) is the time-frequency feature of the abnormal mutation segment, which is obtained by short-time Fourier transform.

3. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 2, characterized in that: The energy distribution fusion equation is specifically expressed as: Where: E(f) is the energy distribution characteristic after fusion; P i (f) is the power spectral density of the i-th feature matrix; w i is the weight coefficient, calculated by entropy weight method; N is the number of feature matrices; η is the weight coefficient of abnormal mutation segment energy; M is the number of abnormal mutation segments; D j is the duration of the jth abnormal mutation segment; T j is the total duration of the signal; E j (f) is the energy spectrum density of the jth abnormal mutation segment.

4. A circuit breaker mechanical fault soundprint feature extraction method according to claim 3, characterized in that: The phase spectrum fusion equation is specifically expressed as: Where: Φ(f) is the phase spectrum feature after fusion; φ i (f) is the phase spectrum of the i-th characteristic matrix; v i is the weight coefficient, obtained through phase consistency evaluation; k is an integer used to ensure phase continuity; μ is the weight coefficient of the abnormal mutation phase; A j is the mutation amplitude of the jth abnormal mutation segment; A max is the maximum mutation amplitude among all abnormal mutation segments; φ j (f) is the phase spectrum of the jth abnormal mutation segment.

5. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 4, characterized in that: The abnormal feature fusion equation is specifically expressed as: Among them: F a (t) is the abnormal feature after fusion; M is the number of abnormal mutation segments; λ j is the weight coefficient of the jth abnormal mutation segment, which is calculated comprehensively by the mutation amplitude and duration; t j,start and t j,end are the start and end time of the jth abnormal mutation segment respectively; t j,mid is the midpoint of the jth abnormal mutation segment; σ j is the time scale parameter of the jth abnormal mutation segment, which is related to the duration; A j is the mutation amplitude of the jth abnormal mutation segment.

6. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 1, characterized in that: The clustering is achieved by using an improved K-means++ algorithm.

7. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 5, characterized in that: The time-frequency characteristics of the abnormal mutation segment are specifically calculated as follows: Among them, s a (t) is the abnormal mutation segment signal, w(t) is the window function, and t represents time.

8. A circuit breaker mechanical fault soundprint feature extraction method according to claim 7, characterized in that: The power spectral density of the characteristic matrix is ​​specifically calculated as follows: Among them, s i (t) is the time domain signal corresponding to the i-th feature matrix, and T is the signal duration.

9. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 8, characterized in that: The energy distribution characteristics after fusion are specifically calculated as follows: Among them, s j (t) is the signal of the jth abnormal mutation segment.

10. A circuit breaker mechanical fault voiceprint feature extraction method according to claim 9, characterized in that: The fused phase spectrum feature is specifically calculated as follows: Among them, X i (f) is the Fourier transform of the i-th feature matrix.

Images