Method and system for identifying mechanical fault of direct-current circuit breaker based on voiceprint feature analysis

Abstract

The invention provides a method and a system for identifying mechanical faults of a direct-current circuit breaker based on voiceprint feature analysis. The method comprises the following steps of: acquiring multi-channel voiceprint signals of the direct-current circuit breaker in a closing process by taking a current signal of an electromagnet coil as a trigger source; performing multilayer wavelet packet decomposition on the multichannel voiceprint signal, extracting a Hilbert envelope energy sequence and sparse codes, and obtaining a test sparse coefficient vector; inputting the test sparse coefficient vector into an adaptive particle swarm optimization support vector machine for state identification to obtain a fault state tag corresponding to the DC circuit breaker; wherein the adaptive particle swarm optimization support vector machine adopts an adaptive particle swarm optimization algorithm to optimize the support vector machine, and fault state classification of the DC circuit breaker is carried out; and the sensitivity, the accuracy and the engineering applicability of mechanical fault identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment fault diagnosis, in particular to a mechanical fault identification method and system of a DC circuit breaker based on voiceprint feature analysis. BACKGROUND

[0002] As a key control and protection device in the power system, the operation state of the DC circuit breaker is directly related to the safety and stability of the device itself and the circuit to which it belongs. Once a mechanical fault occurs, not only will it cause the circuit breaker itself to malfunction or malfunction, but it may also cause cascading accidents such as line protection failure and regional voltage collapse, which will have serious consequences for the safety of the power grid and user power supply. According to the investigation of the International Large Power Grid Conference, mechanical faults account for 80% of circuit breaker faults, and the current breaking failure caused by mechanical faults will directly threaten the stable operation of the entire power grid due to the lack of natural zero-crossing characteristics in the DC system.

[0003] The existing methods for diagnosing mechanical faults of DC circuit breakers mainly focus on analyzing electrical signal features such as current and voltage, or using acceleration and strain signals to achieve state recognition. As a non-contact physical quantity generated by air vibration caused by mechanical action, voiceprint signals have the advantages of convenient signal acquisition and sensitive response, and have gradually become a new information source for equipment fault diagnosis. However, due to the short-time, non-stationary, and nonlinear characteristics of the voiceprint signals of the circuit breaker, traditional time-domain, frequency-domain, or single-scale analysis methods cannot effectively capture the voiceprint changes caused by minor faults, and the feature extraction capability is limited. Moreover, the existing classification and recognition methods rely on fixed parameter support vector machines or shallow neural networks, which have poor adaptability to changes in fault sample distribution and are difficult to balance accuracy and generalization ability. SUMMARY

[0004] In order to solve the problems of signal acquisition difficulty, insufficient feature extraction, low classification accuracy, and poor model generalization ability in the existing technology of circuit breaker mechanical fault recognition, the present application provides a DC circuit breaker mechanical fault recognition method based on voiceprint feature analysis, which improves in that the method comprises:

[0005] Using the electromagnet coil current signal as a trigger source to collect multi-channel voiceprint signals of the DC circuit breaker during the closing process;

[0006] Performing multi-layer wavelet packet decomposition on the multi-channel voiceprint signals to extract Hilbert envelope energy sequences and sparse codes, and obtaining a test sparse coefficient vector;

[0007] Inputting the test sparse coefficient vector into an adaptive particle swarm optimization support vector machine for state recognition to obtain a fault state label corresponding to the DC circuit breaker;

[0008] The adaptive particle swarm optimization support vector machine adopts an adaptive particle swarm optimization algorithm to optimize the support vector machine to classify the fault states of the DC circuit breaker.

[0009] Preferably, the training process of the adaptive particle swarm optimization support vector machine comprises:

[0010] A plurality of states of the DC circuit breaker are simulated, the states including normal, electromagnetic drive hysteresis fault, contact wear fault and buffer failure, and voiceprint signals of a predetermined number of the plurality of states are synchronously collected;

[0011] Feature extraction is performed on the voiceprint signals of the plurality of states to obtain training sparse coefficient vectors;

[0012] The training sparse coefficient vectors are taken as input data, corresponding state labels are taken as output data, and the input data and the output data are taken as a training sample set to input a support vector machine;

[0013] A radial basis kernel function is selected as a kernel function form of the support vector machine, and model hyperparameters and a maximum number of iterations of the support vector machine are initialized;

[0014] The initialized model hyperparameters are taken as position vectors of particles, positions and speeds of the particle swarm are updated, and the best model hyperparameters are obtained when a fitness function of the particles reaches a precision threshold;

[0015] The training sample set is divided into a training set and a validation set under the best model hyperparameters to train a support vector machine model and validate the support vector machine model;

[0016] The training is terminated when the maximum number of iterations is reached or a target accuracy is reached, and a trained adaptive particle swarm optimization support vector machine is obtained.

[0017] Preferably, the feature extraction on the voiceprint signals of the plurality of states to obtain the training sparse coefficient vectors comprises:

[0018] The voiceprint signals of the plurality of states are subjected to multi-layer wavelet packet decomposition, Hilbert envelope energy sequences are extracted, and wavelet packet envelope spectrum entropy features are obtained;

[0019] The wavelet packet envelope spectrum entropy features are used to construct training feature vectors of the plurality of states at the same time, and a predetermined number of the training feature vectors are combined to form a training feature matrix;

[0020] The training feature matrix is trained and solved by a sparse dictionary learning algorithm to obtain the training sparse coefficient vectors.

[0021] Preferably, the training of the feature matrix by the sparse dictionary learning algorithm and the solving of the training sparse coefficient vectors comprise:

[0022] initializing a dictionary D;

[0023] solving sparse coding with respect to the training feature matrix using an orthogonal matching pursuit algorithm to obtain a training sparse coefficient vector;

[0024] updating the dictionary using singular value decomposition to minimize the reconstruction error;

[0025] alternately performing sparse coding and dictionary updating until a maximum allowed number of non-zero elements is not greater than a control sparsity parameter T0, convergence is completed, and a final sparse dictionary and training sparse coefficient vector are obtained.

[0026] Preferably, the multi-channel voiceprint signal is subjected to multi-layer wavelet packet decomposition, and a Hilbert envelope energy sequence and sparse coding are extracted to obtain a test sparse coefficient vector, including:

[0027] The multi-channel voiceprint signal is subjected to multi-layer wavelet packet decomposition using a Symlet wavelet basis function;

[0028] The multi-channel voiceprint signal subjected to wavelet packet decomposition is subjected to Hilbert transform to extract a Hilbert envelope energy sequence to obtain a test feature vector;

[0029] The test feature vector is subjected to sparse coding on the sparse dictionary to obtain a test sparse coefficient vector.

[0030] Preferably, after the multi-channel voiceprint signal of the DC circuit breaker during the closing process is collected using the electromagnetic coil current signal as a trigger source, before the multi-channel voiceprint signal is subjected to multi-layer wavelet packet decomposition, the method further includes:

[0031] The collected voiceprint signal is preprocessed.

[0032] Preferably, the preprocessing of the collected voiceprint signal includes:

[0033] The continuous-time signal is sampled by an analog-to-digital converter to obtain a discrete-time signal;

[0034] The discrete-time signal is subjected to a DC offset removal operation;

[0035] The discrete-time signal is subjected to wavelet decomposition into wavelet coefficients of different scales, the wavelet coefficients of different scales are denoised using a soft threshold function, and the discrete-time signal is reconstructed using the denoised wavelet coefficients;

[0036] The reconstructed discrete-time signal is subjected to short-time analysis using a fixed frame segmentation method with a frame length of N, and a Hamming window function is introduced for window processing.

[0037] Based on the same inventive concept, the application further provides a DC circuit breaker mechanical fault recognition system based on voiceprint feature analysis, including:

[0038] a collection module, configured to collect multi-channel voiceprint signals of the DC circuit breaker in a closing process with an electromagnet coil current signal as a trigger source;

[0039] a feature extraction module, configured to perform multi-layer wavelet packet decomposition on the multi-channel voiceprint signals, extract a Hilbert envelope energy sequence and sparse coding, and obtain a test sparse coefficient vector;

[0040] a state identification module, configured to input the test sparse coefficient vector into an adaptive particle swarm optimization support vector machine to perform state identification, and obtain a fault state label corresponding to the DC circuit breaker;

[0041] The adaptive particle swarm optimization support vector machine is used to optimize the support vector machine by using an adaptive particle swarm optimization algorithm to classify the fault states of the DC circuit breaker.

[0042] Preferably, the training process of the adaptive particle swarm optimization support vector machine in the state identification module includes:

[0043] a plurality of states of the DC circuit breaker are simulated, the states including normal, electromagnetic drive hysteresis fault, contact wear fault and buffer failure, and voiceprint signals of a predetermined number of the plurality of states are synchronously collected;

[0044] feature extraction is performed on the voiceprint signals of the plurality of states to obtain a training sparse coefficient vector;

[0045] the training sparse coefficient vector is taken as input data, a corresponding state label is taken as output data, and the input data and the output data are taken as a training sample set to input a support vector machine;

[0046] a radial basis kernel function is selected as a kernel function form of the support vector machine, and model hyperparameters and a maximum number of iterations of the support vector machine are initialized;

[0047] the initialized model hyperparameters are taken as position vectors of particles, positions and speeds of the particle swarm are updated, and the best model hyperparameters are obtained when a fitness function of the particles reaches a precision threshold;

[0048] the training sample set is divided into a training set and a verification set under the best model hyperparameters to train a support vector machine model and verify the support vector machine model;

[0049] the training is terminated when the maximum number of iterations is reached or a target accuracy is reached, and a trained adaptive particle swarm optimization support vector machine is obtained.

[0050] Preferably, the feature extraction performed on the voiceprint signals of the plurality of states to obtain the training sparse coefficient vector in the state identification module includes:

[0051] The multiple states of the voiceprint signal are subjected to multi-layer wavelet packet decomposition, a Hilbert envelope energy sequence is extracted, and a wavelet packet envelope spectrum entropy feature is obtained.

[0052] The wavelet packet envelope spectrum entropy features are used to construct training feature vectors of the multiple states at the same time, and a predetermined number of training feature vectors are combined to form a training feature matrix.

[0053] The training feature matrix is trained by a sparse dictionary learning algorithm, and a training sparse coefficient vector is obtained.

[0054] Preferably, the state recognition module trains the feature matrix by a sparse dictionary learning algorithm, and a training sparse coefficient vector is obtained, comprising:

[0055] Initializing a dictionary D;

[0056] The sparse coding of the training feature matrix is solved by using an orthogonal matching pursuit algorithm, and a training sparse coefficient vector is obtained;

[0057] The dictionary is updated by singular value decomposition to minimize the reconstruction error;

[0058] The sparse coding and the dictionary updating are alternately performed until the maximum number of non-zero elements is not greater than a control sparsity parameter T0, and convergence is completed, and a final sparse dictionary and a training sparse coefficient vector are obtained.

[0059] Preferably, the feature extraction module subjects the multi-channel voiceprint signal to multi-layer wavelet packet decomposition, extracts a Hilbert envelope energy sequence and sparse coding, and obtains a test sparse coefficient vector, comprising:

[0060] The multi-channel voiceprint signal is subjected to multi-layer wavelet packet decomposition by using a Symlet wavelet basis function;

[0061] The multi-channel voiceprint signal subjected to wavelet packet decomposition is subjected to Hilbert transform, a Hilbert envelope energy sequence is extracted, and a test feature vector is obtained;

[0062] The test feature vector is subjected to sparse coding on the sparse dictionary to obtain a test sparse coefficient vector.

[0063] Preferably, the system further comprises a preprocessing module for preprocessing the collected voiceprint signal.

[0064] Preferably, the preprocessing module is specifically configured to:

[0065] The continuous-time signal is sampled by an analog-to-digital converter to obtain a discrete-time signal;

[0066] The discrete-time signal is subjected to a direct current offset operation;

[0067] The discrete-time signal is decomposed into wavelet coefficients of different scales, and the wavelet coefficients of different scales are denoised using a soft thresholding function. The discrete-time signal is then reconstructed using the denoised wavelet coefficients.

[0068] A fixed-frame segmentation method with a frame length of N is used to perform short-time analysis on the reconstructed discrete-time signal, and a Hamming window function is introduced for windowing processing.

[0069] Furthermore, this application also provides a computing device, comprising: at least one processor and a memory;

[0070] The memory is used to store one or more programs;

[0071] When the one or more programs are executed by the one or more processors, a method for identifying mechanical faults in DC circuit breakers based on voiceprint feature analysis, as described above, is implemented.

[0072] In another aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-described method for identifying mechanical faults in DC circuit breakers based on voiceprint feature analysis.

[0073] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0074] This invention provides a method and system for identifying mechanical faults in DC circuit breakers based on acoustic signature feature analysis. The method includes: acquiring multi-channel acoustic signature signals of the DC circuit breaker during the closing process using an electromagnet coil current signal as a trigger source; performing multi-level wavelet packet decomposition on the multi-channel acoustic signature signals to extract the Hilbert envelope energy sequence and sparse coding, obtaining a test sparse coefficient vector; inputting the test sparse coefficient vector into an adaptive particle swarm optimization support vector machine for state identification, obtaining the corresponding fault state label of the DC circuit breaker; wherein, the adaptive particle swarm optimization support vector machine uses an adaptive particle swarm optimization algorithm to optimize the support vector machine and classify the fault state of the DC circuit breaker. The method and system for identifying mechanical faults in DC circuit breakers based on acoustic signature feature analysis provided by this invention improves the sensitivity, accuracy, and engineering applicability of mechanical fault identification. By constructing a multi-scale frequency-energy-time three-dimensional feature map of the acoustic signature signal, combined with sparse representation and an optimized classifier strategy, it achieves sensitive identification of three typical mechanical faults in DC circuit breakers, providing a non-contact intelligent diagnostic solution for key equipment in power systems. Attached Figure Description

[0075] Figure 1 A flowchart of a method for identifying mechanical faults in DC circuit breakers based on acoustic signature analysis provided by the present invention;

[0076] Figure 2The present invention provides a flowchart for identifying mechanical faults in DC circuit breakers.

[0077] Figure 3 The sparse dictionary learning flowchart provided by this invention;

[0078] Figure 4 The APSO optimized SVM flowchart provided by this invention;

[0079] Figure 5 The optimized results of fault feature identification provided by this invention;

[0080] Figure 6 A structural diagram of a DC circuit breaker mechanical fault identification system based on acoustic signature analysis provided by the present invention;

[0081] Figure 7 An electronic device provided by the present invention. Detailed Implementation

[0082] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0083] Example 1

[0084] This invention provides a method for identifying mechanical faults in DC circuit breakers based on acoustic signature analysis, such as... Figure 1 ,include:

[0085] The multi-channel acoustic fingerprint signal of the DC circuit breaker during the closing process is acquired using the current signal of the electromagnet coil as a trigger source.

[0086] Multi-level wavelet packet decomposition was performed on the multi-channel acoustic fingerprint signal to extract the Hilbert envelope energy sequence and sparse coding, and the test sparse coefficient vector was obtained.

[0087] The test sparse coefficient vector is input into an adaptive particle swarm optimization support vector machine for state identification, and the fault state label corresponding to the DC circuit breaker is obtained.

[0088] Among them, the adaptive particle swarm optimization support vector machine uses the adaptive particle swarm optimization algorithm to optimize the support vector machine and classify the fault states of DC circuit breakers.

[0089] Preferably, the training process of the adaptive particle swarm optimization support vector machine includes:

[0090] Simulate multiple states of a DC circuit breaker, including normal, electromagnetic drive hysteresis fault, contact wear fault, and buffer failure, and synchronously collect a predetermined number of acoustic fingerprint signals for these multiple states.

[0091] Feature extraction is performed on the voiceprint signals of the multiple states to obtain a training sparse coefficient vector;

[0092] The training sparse coefficient vector is used as input data, the corresponding state label is used as output data, and the input data and output data are used as training sample sets to input the support vector machine.

[0093] The radial basis function kernel is selected as the kernel function form of the support vector machine, and the model hyperparameters and maximum number of iterations of the support vector machine are initialized.

[0094] The initial model hyperparameters are used as the position vectors of the particles to update the position and velocity of the particle swarm. The optimal model hyperparameters are obtained when the fitness function of the particles reaches the accuracy threshold.

[0095] Under the optimal model hyperparameters, the training sample set is divided into a training set and a validation set for training and validating the support vector machine model.

[0096] Training terminates when the maximum number of iterations is reached or the target accuracy is achieved, resulting in a well-trained adaptive particle swarm optimization support vector machine.

[0097] Preferably, the step of extracting features from the voiceprint signals of the multiple states to obtain a training sparse coefficient vector includes:

[0098] Multi-level wavelet packet decomposition is performed on the voiceprint signals of the multiple states to extract the Hilbert envelope energy sequence and obtain the wavelet packet envelope spectral entropy features;

[0099] Training feature vectors for multiple states at the same time are constructed using wavelet envelope spectrum entropy features, and a predetermined number of training feature vectors are combined to form a training feature matrix.

[0100] The training feature matrix is ​​trained and the training sparse coefficient vector is obtained by solving the sparse dictionary learning algorithm.

[0101] Preferably, the step of training the feature matrix and obtaining the training sparse coefficient vector through a sparse dictionary learning algorithm includes:

[0102] Initialize dictionary D;

[0103] The orthogonal matching pursuit algorithm is used to solve the sparse coding of the training feature matrix, and the training sparse coefficient vector is obtained.

[0104] Updating the dictionary using singular value decomposition minimizes the reconstruction error;

[0105] The process of solving the sparse encoding and updating the dictionary is performed alternately until the maximum number of allowed non-zero elements is no greater than the control sparsity parameter T0, at which point convergence is achieved, and the final sparse dictionary and training sparse coefficient vector are obtained.

[0106] Preferably, the step of performing multi-level wavelet packet decomposition on the multi-channel voiceprint signal to extract the Hilbert envelope energy sequence and sparse coding to obtain the test sparse coefficient vector includes:

[0107] Multi-level wavelet packet decomposition was performed on the multi-channel voiceprint signal using Symlet wavelet basis functions.

[0108] The Hilbert transform is performed on the multi-channel acoustic signal after wavelet packet decomposition, and the Hilbert envelope energy sequence is extracted to obtain the test feature vector;

[0109] The test feature vector is sparsely encoded on a sparse dictionary to obtain the test sparse coefficient vector.

[0110] Preferably, after acquiring the multi-channel acoustic signature signal of the DC circuit breaker during the closing process using the electromagnet coil current signal as a trigger source, and before performing multi-layer wavelet packet decomposition on the multi-channel acoustic signature signal, the method further includes:

[0111] The collected voiceprint signals are preprocessed.

[0112] Preferably, the preprocessing of the acquired voiceprint signal includes:

[0113] A discrete-time signal is obtained by sampling a continuous-time signal through an analog-to-digital converter.

[0114] Perform DC offset removal operation on discrete-time signals;

[0115] The discrete-time signal is decomposed into wavelet coefficients of different scales, and the wavelet coefficients of different scales are denoised using a soft thresholding function. The discrete-time signal is then reconstructed using the denoised wavelet coefficients.

[0116] A fixed-frame segmentation method with a frame length of N is used to perform short-time analysis on the reconstructed discrete-time signal, and a Hamming window function is introduced for windowing processing.

[0117] Example 2

[0118] Based on the same inventive concept, this invention also provides a preferred embodiment of a method for identifying mechanical faults in DC circuit breakers based on voiceprint feature analysis, such as... Figure 2 ,include:

[0119] This application takes a hybrid DC circuit breaker as an example. Considering the large number of mechanical components involved in its operation during the closing process, the violent action, and the accompanying obvious acoustic response, the acoustic fingerprint signal during the closing phase is selected for analysis. Acoustic fingerprint signals during the closing phase of the circuit breaker are collected under normal operating conditions and three artificially set typical fault conditions.

[0120] Step 1: Since multi-scale wavelet packet envelope spectrum is sensitive to non-stationary transient impact signals and sparse dictionary learning can extract structural acoustic patterns, three types of faults covering the entire mechanical action process are selected for simulation: electromagnetic drive hysteresis, contact wear, and buffer failure.

[0121] Electromagnetic drive hysteresis fault is simulated by inserting a small foreign object into the gap of the electromagnet core movement to simulate the operation state of core jamming or drive delay. The acoustic signal is manifested as delayed transient response and delayed local impact in the initial stage of closing.

[0122] Contact wear failure is caused by artificially creating uneven wear on the contact surface of the contact, simulating the physical wear or surface contamination state of the contact after long-term operation, which leads to secondary rebound or poor contact during the closing process, resulting in multiple local impact events or bi-peak signal structure.

[0123] Buffer failure can be simulated by removing the mechanical buffer device or replacing it with rigid material to mimic the aging and shedding of rubber damping, causing direct collision of structural components and generating high-amplitude transient impact noise.

[0124] Four MEMS microphones are evenly distributed outside the housing of the DC circuit breaker's mechanical switching unit, with a sampling frequency set to 96kHz. An NI USB-4431 data acquisition card is used for synchronous acquisition of audio signals, and the data is transmitted in real-time to a local server for storage and post-processing via a USB interface. The excitation current signal of the circuit breaker's closing coil is used as the trigger source, and a hardware triggering mechanism initiates the acquisition process, thereby standardizing the start time reference for the voiceprint samples.

[0125] Step 2: To ensure signal quality and feature extraction effectiveness, the original voiceprint signal undergoes multi-stage preprocessing after acquisition to suppress low-frequency drift, background noise, and abrupt interference. The same preprocessing method is used in both the training and testing phases.

[0126] The acquired raw speaker signals were sequentially processed using DC offset removal, wavelet soft thresholding denoising, and frame windowing to remove background noise, suppress spectral leakage, and enhance the extraction capability of short-time features. The noise standard deviation was estimated from the median of the highest frequency wavelet coefficients, and the threshold was dynamically calculated using an adaptive formula to ensure the robustness of the denoising process.

[0127] The continuous-time signal x(t) is sampled by an analog-to-digital converter to obtain the discrete-time signal x(n).

[0128] Multi-scale wavelet soft thresholding denoising is employed. Wavelet decomposition is performed on x(n) to obtain a multi-level wavelet coefficient sequence. Then, a soft thresholding compression function is applied to the detail coefficients of each frequency band, as shown in the following expression:

[0129]

[0130] Where, ω i Let λ be the wavelet coefficient sequence, and λ be the adaptive threshold in the wavelet soft thresholding function.

[0131] The preprocessed discrete voiceprint signal x(n) is segmented into fixed frames of length N for short-time analysis. A Hamming window function is introduced to reduce boundary effects and enhance local feature extraction capabilities. The windowing expression is as follows:

[0132]

[0133] Where, x ω (n) represents the signal after windowing.

[0134] Step 3: Perform multi-level wavelet packet decomposition on the preprocessed signal, extract the Hilbert envelope energy sequence of each sub-band signal, and construct a three-dimensional feature map of frequency-energy-time; normalize the envelope energy sequence of each band and calculate its information entropy to form a discriminative envelope spectral entropy feature vector, which is used to compress and represent the multi-frequency non-stationary features during the circuit breaker operation process.

[0135] The preprocessed discrete voiceprint signal is decomposed into L-level wavelet packets using Symlet wavelet basis functions to obtain 2 L Sub-band signal x i (n), where i = 1, 2, ..., 2 L Let i represent the i-th sub-band signal. Perform a Hilbert transform on each sub-band signal to extract its envelope signal, obtaining the instantaneous energy envelope sequence representing the signal in sub-band i:

[0136] E i (n)=|H[x i (n)]|

[0137] Among them, E i (n) is the instantaneous energy envelope sequence of the signal in sub-band i.

[0138] Normalize the energy sequence for each frequency band and calculate its envelope spectral entropy H. i :

[0139]

[0140] Among them, E i (k) represents the k-th instantaneous energy of the signal in sub-band i, p i,k Let be the relative proportion of the energy of the k-th frame in the i-th sub-band signal within that frequency band.

[0141]

[0142] Finally, the feature vector H is constructed:

[0143] H = [H1,H2,...,H] 2L ]

[0144] Step 4: During each circuit breaker closing operation, four MEMS microphones simultaneously acquire voiceprint signals, and extract their wavelet packet envelope spectral entropy features H. (1) H (2) H (3) H (4) ∈R m To improve the completeness of the overall fault description, a feature concatenation method is used to construct a fused feature vector y for each sample. j =[H (1) H (2) H (3) H (4) ]∈R 4m×n y j Let m be the feature vector of the j-th sample, and m is 2. L 4m represents four paths, n is the total number of samples, indicating how many sets of data were collected, and all samples form the feature matrix Y = [y1, y2, ..., y]. n ]∈R 4m×n y n Let be the feature vector of the nth sample.

[0145] Step 5: Use the K-SVD sparse dictionary learning algorithm to sparsely model the envelope spectral entropy features of the training samples, construct a feature dictionary with strong representativeness and high expressive power, and project the test samples onto the dictionary through sparse coding methods such as orthogonal matching pursuit to obtain structural sparse coefficients as input features for fault identification.

[0146] The feature matrix Y is trained using the K-SVD algorithm learned through a sparse dictionary, such as... Figure 3 The Orthogonal Matching Pursuit (OMP) algorithm is used to solve the sparse coding of the training feature matrix, and the Singular Value Decomposition (SVD) algorithm is used to iteratively update the dictionary elements to obtain the dictionary matrix D∈R. 4d×K With the sparse coefficient matrix X∈R K×n d represents the feature dimension of each signal after dimensionality reduction, K represents the number of feature bases in the sparse dictionary, T0 controls the sparsity, and the Frobenius norm measures the reconstruction error. The optimization objective function is as follows:

[0147]

[0148] Where T0 is a hyperparameter controlling sparsity, representing the value of each x. i The maximum number of non-zero elements allowed in x, where F is the norm, and x jLet be the final sparse coefficient vector of the j-th sample.

[0149] The new test sample y to be identified j Perform sparse encoding on the trained dictionary D and solve for its sparse coefficient vector:

[0150]

[0151] Where x is the sparse coefficient vector of the new input sample in dictionary D.

[0152] Step 6: Construct a Support Vector Machine (SVM) classification model, selecting a radial basis function kernel, and train it based on sparse feature input. To improve model performance and avoid reliance on manual parameter setting, an adaptive particle swarm optimization (PSO) algorithm is introduced to globally optimize the hyperparameters C and kernel function parameter γ of the SVM. The cross-validation accuracy is used as the fitness function, and the speed and position are dynamically adjusted through particle swarm iteration to find the optimal solution. Finally, the sparse feature vectors are input into the optimized SVM model to achieve accurate identification of the fault type of the test samples.

[0153] Take 50 sets of voiceprint signals each from normal and three simulated fault states, extract the wavelet envelope spectrum entropy features of each set of data, and then use the sparse coefficient vector x of all training samples obtained from sparse dictionary learning. j ∈R K Its corresponding fault status label y j The training dataset T = {x ∈ {1,2,3,4} is composed of ∈ {1,2,3,4}. j ,y j Adaptive Particle Swarm Optimization Support Vector Machine (APSO-SVM) is used for state identification.

[0154] Choose the radial basis kernel function K(x) j (x) is the kernel function form of Support Vector Machine (SVM). The model hyperparameters include the penalty factor C and the kernel width parameter γ. The SVM model is constructed as follows:

[0155]

[0156] K(x j ,x)=exp(-γ||x j -x|| 2 )

[0157] Where f(x) is the expression for the support vector machine, α j Let b be the Lagrange multiplier of the support vector machine, and b be the bias term of the decision function.

[0158] The support vector machine parameters (C, γ) to be optimized are represented as particle position vectors, and the adaptive particle swarm optimization algorithm (APSO) is used for optimization.Figure 4 By introducing dynamic inertia weight ω control, the particle's fitness function is set as the recognition accuracy based on cross-validation. The particle velocity and position update formulas are as follows:

[0159]

[0160] Where ω(t) is the particle's inertial weight, Let be the velocity of the i-th particle in the (t+1)-th iteration. Let c1 be the velocity of the i-th particle in the t-th iteration, c2 be the individual learning factor, c2 be the group learning factor, and r1 and r2 be random numbers in the range [0, 1]. i Let g be the individual optimal position of the i-th particle, and g be the global optimal position. Let i be the position of the i-th particle in the (t+1)-th iteration. Let be the position of the i-th particle in the t-th iteration.

[0161] Based on the state identification results, such as Figure 5 It can be seen that by constructing a multi-scale frequency-energy-time three-dimensional feature map of the acoustic signal, and combining sparse representation and optimized classifier strategy, sensitive identification of three typical mechanical faults of DC circuit breakers is achieved, providing a non-contact intelligent diagnostic solution for key equipment in the power system.

[0162] Example 3

[0163] Based on the same inventive concept, this invention also provides a DC circuit breaker mechanical fault identification system based on voiceprint feature analysis, such as... Figure 6 ,include:

[0164] The acquisition module is used to acquire multi-channel acoustic fingerprint signals of a DC circuit breaker during the closing process, using the electromagnet coil current signal as a trigger source.

[0165] The feature extraction module is used to perform multi-level wavelet packet decomposition on multi-channel voiceprint signals, extract the Hilbert envelope energy sequence and sparse coding, and obtain the test sparse coefficient vector.

[0166] The state identification module is used to input the test sparse coefficient vector into the adaptive particle swarm optimization support vector machine for state identification, and obtain the fault state label corresponding to the DC circuit breaker.

[0167] Among them, the adaptive particle swarm optimization support vector machine uses the adaptive particle swarm optimization algorithm to optimize the support vector machine and classify the fault states of DC circuit breakers.

[0168] Preferably, the training process of the adaptive particle swarm optimization support vector machine in the state identification module includes:

[0169] Simulate multiple states of a DC circuit breaker, including normal, electromagnetic drive hysteresis fault, contact wear fault, and buffer failure, and synchronously collect a predetermined number of acoustic fingerprint signals for these multiple states.

[0170] Feature extraction is performed on the voiceprint signals of the multiple states to obtain a training sparse coefficient vector;

[0171] The training sparse coefficient vector is used as input data, the corresponding state label is used as output data, and the input data and output data are used as training sample sets to input the support vector machine.

[0172] The radial basis function kernel is selected as the kernel function form of the support vector machine, and the model hyperparameters and maximum number of iterations of the support vector machine are initialized.

[0173] The initial model hyperparameters are used as the position vectors of the particles to update the position and velocity of the particle swarm. The optimal model hyperparameters are obtained when the fitness function of the particles reaches the accuracy threshold.

[0174] Under the optimal model hyperparameters, the training sample set is divided into a training set and a validation set for training and validating the support vector machine model.

[0175] Training terminates when the maximum number of iterations is reached or the target accuracy is achieved, resulting in a well-trained adaptive particle swarm optimization support vector machine.

[0176] Preferably, the state identification module performs feature extraction on the voiceprint signals of the multiple states to obtain a training sparse coefficient vector, including:

[0177] Multi-level wavelet packet decomposition is performed on the voiceprint signals of the multiple states to extract the Hilbert envelope energy sequence and obtain the wavelet packet envelope spectral entropy features;

[0178] Training feature vectors for multiple states at the same time are constructed using wavelet envelope spectrum entropy features, and a predetermined number of training feature vectors are combined to form a training feature matrix.

[0179] The training feature matrix is ​​trained and the training sparse coefficient vector is obtained by solving the sparse dictionary learning algorithm.

[0180] Preferably, the state identification module trains the feature matrix and solves for the training sparse coefficient vector using a sparse dictionary learning algorithm, including:

[0181] Initialize dictionary D;

[0182] The orthogonal matching pursuit algorithm is used to solve the sparse coding of the training feature matrix, and the training sparse coefficient vector is obtained.

[0183] Updating the dictionary using singular value decomposition minimizes the reconstruction error;

[0184] The process of solving the sparse encoding and updating the dictionary is performed alternately until the maximum number of allowed non-zero elements is no greater than the control sparsity parameter T0, at which point convergence is achieved, and the final sparse dictionary and training sparse coefficient vector are obtained.

[0185] Preferably, the feature extraction module performs multi-level wavelet packet decomposition on the multi-channel voiceprint signal to extract the Hilbert envelope energy sequence and sparse coding, obtaining a test sparse coefficient vector, including:

[0186] Multi-level wavelet packet decomposition was performed on the multi-channel voiceprint signal using Symlet wavelet basis functions.

[0187] The Hilbert transform is performed on the multi-channel acoustic signal after wavelet packet decomposition, and the Hilbert envelope energy sequence is extracted to obtain the test feature vector;

[0188] The test feature vector is sparsely encoded on a sparse dictionary to obtain the test sparse coefficient vector.

[0189] Preferably, the system further includes a preprocessing module for preprocessing the acquired voiceprint signals.

[0190] Preferably, the preprocessing module is specifically used for:

[0191] A discrete-time signal is obtained by sampling a continuous-time signal through an analog-to-digital converter.

[0192] Perform DC offset removal operation on discrete-time signals;

[0193] The discrete-time signal is decomposed into wavelet coefficients of different scales, and the wavelet coefficients of different scales are denoised using a soft thresholding function. The discrete-time signal is then reconstructed using the denoised wavelet coefficients.

[0194] A fixed-frame segmentation method with a frame length of N is used to perform short-time analysis on the reconstructed discrete-time signal, and a Hamming window function is introduced for windowing processing.

[0195] Example 4

[0196] Based on the same inventive concept, the present invention also provides an electronic device, such as... Figure 7As shown, the electronic device may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0197] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the DC circuit breaker mechanical fault identification method based on voiceprint feature analysis in the above embodiments.

[0198] Example 5

[0199] Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the DC circuit breaker mechanical fault identification method based on voiceprint feature analysis in the above embodiments.

[0200] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0201] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0202] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0203] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0204] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval. The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

Claims

1. A method for identifying mechanical faults in DC circuit breakers based on acoustic signature analysis, characterized in that, include: The multi-channel acoustic fingerprint signal of the DC circuit breaker during the closing process is acquired using the current signal of the electromagnet coil as a trigger source. Multi-level wavelet packet decomposition was performed on the multi-channel acoustic fingerprint signal to extract the Hilbert envelope energy sequence and sparse coding, and the test sparse coefficient vector was obtained. The test sparse coefficient vector is input into an adaptive particle swarm optimization support vector machine for state identification, and the fault state label corresponding to the DC circuit breaker is obtained. Among them, the adaptive particle swarm optimization support vector machine uses the adaptive particle swarm optimization algorithm to optimize the support vector machine and classify the fault states of DC circuit breakers.

2. The method as described in claim 1, characterized in that, The training process of the adaptive particle swarm optimization support vector machine includes: Simulate multiple states of a DC circuit breaker, including normal, electromagnetic drive hysteresis fault, contact wear fault, and buffer failure, and synchronously collect a predetermined number of acoustic fingerprint signals for these multiple states. Feature extraction is performed on the voiceprint signals of the multiple states to obtain a training sparse coefficient vector; The training sparse coefficient vector is used as input data, the corresponding state label is used as output data, and the input data and output data are used as training sample sets to input the support vector machine. The radial basis function kernel is selected as the kernel function form of the support vector machine, and the model hyperparameters and maximum number of iterations of the support vector machine are initialized. The initial model hyperparameters are used as the position vectors of the particles to update the position and velocity of the particle swarm. The optimal model hyperparameters are obtained when the fitness function of the particles reaches the accuracy threshold. Under the optimal model hyperparameters, the training sample set is divided into a training set and a validation set for training and validating the support vector machine model. Training terminates when the maximum number of iterations is reached or the target accuracy is achieved, resulting in a well-trained adaptive particle swarm optimization support vector machine.

3. The method as described in claim 2, characterized in that, The step of extracting features from the voiceprint signals of the multiple states to obtain a training sparse coefficient vector includes: Multi-level wavelet packet decomposition is performed on the voiceprint signals of the multiple states to extract the Hilbert envelope energy sequence and obtain the wavelet packet envelope spectral entropy features; Training feature vectors for multiple states at the same time are constructed using wavelet envelope spectrum entropy features, and a predetermined number of training feature vectors are combined to form a training feature matrix. The training feature matrix is ​​trained and the training sparse coefficient vector is obtained by solving the sparse dictionary learning algorithm.

4. A method as described in claim 3, characterized in that, The step of training the feature matrix and solving for the training sparse coefficient vector using a sparse dictionary learning algorithm includes: Initialize dictionary D; The orthogonal matching pursuit algorithm is used to solve the sparse coding of the training feature matrix, and the training sparse coefficient vector is obtained. Updating the dictionary using singular value decomposition minimizes the reconstruction error; The process of solving the sparse encoding and updating the dictionary is performed alternately until the maximum number of allowed non-zero elements is no greater than the control sparsity parameter T0, at which point convergence is achieved, and the final sparse dictionary and training sparse coefficient vector are obtained.

5. A method as described in claim 4, characterized in that, The process involves performing multi-level wavelet packet decomposition on the multi-channel acoustic signal to extract the Hilbert envelope energy sequence and sparse coding, resulting in a test sparse coefficient vector, including: Multi-level wavelet packet decomposition was performed on the multi-channel voiceprint signal using Symlet wavelet basis functions. The Hilbert transform is performed on the multi-channel acoustic signal after wavelet packet decomposition, and the Hilbert envelope energy sequence is extracted to obtain the test feature vector; The test feature vector is sparsely encoded on a sparse dictionary to obtain the test sparse coefficient vector.

6. A method as described in claim 1, characterized in that, After acquiring the multi-channel acoustic signature signal of the DC circuit breaker during the closing process using the electromagnet coil current signal as a trigger source, and before performing multi-layer wavelet packet decomposition on the multi-channel acoustic signature signal, the method further includes: The collected voiceprint signals are preprocessed.

7. A method as described in claim 6, characterized in that, The preprocessing of the acquired voiceprint signals includes: A discrete-time signal is obtained by sampling a continuous-time signal through an analog-to-digital converter. Perform DC offset removal operation on discrete-time signals; The discrete-time signal is decomposed into wavelet coefficients of different scales, and the wavelet coefficients of different scales are denoised using a soft thresholding function. The discrete-time signal is then reconstructed using the denoised wavelet coefficients. A fixed-frame segmentation method with a frame length of N is used to perform short-time analysis on the reconstructed discrete-time signal, and a Hamming window function is introduced for windowing processing.

8. A DC circuit breaker mechanical fault identification system based on voiceprint feature analysis, characterized in that, include: The acquisition module is used to acquire multi-channel acoustic fingerprint signals of a DC circuit breaker during the closing process, using the electromagnet coil current signal as a trigger source. The feature extraction module is used to perform multi-level wavelet packet decomposition on multi-channel voiceprint signals, extract the Hilbert envelope energy sequence and sparse coding, and obtain the test sparse coefficient vector. The state identification module is used to input the test sparse coefficient vector into the adaptive particle swarm optimization support vector machine for state identification, and obtain the fault state label corresponding to the DC circuit breaker. Among them, the adaptive particle swarm optimization support vector machine uses the adaptive particle swarm optimization algorithm to optimize the support vector machine and classify the fault states of DC circuit breakers.

9. A computer device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for identifying mechanical faults in DC circuit breakers based on acoustic signature analysis as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that, It contains an execution program, which, when executed, implements a method for identifying mechanical faults in DC circuit breakers based on voiceprint feature analysis as described in any one of claims 1 to 7.

Images