Electrical characteristic signal extraction method of power equipment in complex working condition environment

Abstract

The invention provides a method for extracting electrical characteristic signals of electrical equipment in a complex working condition environment, and belongs to the technical field of electrical equipment detection.The method comprises the steps that a multi-channel ultrasonic sensor array is arranged to collect partial discharge signals, background noise is eliminated through adaptive noise cancellation processing, and an ultra-sparse frequency band energy distribution vector is constructed; the method comprises the following steps: calling a self-adaptive time-frequency analysis model to extract instantaneous frequency, amplitude and phase parameters to form a micro-hour-frequency characteristic matrix, separating independent source signals through independent component analysis, calculating a kurtosis value and a skewness value, fusing multi-domain characteristics to construct a transient stationary comprehensive characteristic vector, matching with a standard discharge characteristic vector library to identify the discharge type and intensity, and calculating the discharge intensity. A corresponding prediction algorithm is selected according to the discharge mode, multi-parameter coupling optimization adjustment is started under a certain condition, an electrical characteristic signal description vector is finally constructed, and the technical problem that the partial discharge signal of the power equipment is difficult to accurately extract under a complex working condition environment is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power equipment detection, and in particular relates to a method for extracting electrical characteristic signals of power equipment under complex working conditions. BACKGROUND

[0002] The extraction of partial discharge signals of power equipment is an important technical means for monitoring the state of power systems. Traditional methods mainly use a single sensor in combination with fixed threshold detection or simple filtering algorithms to extract discharge signals. However, in the complex working condition environment of actual power equipment operation, there are various interference factors such as strong electromagnetic interference, mechanical vibration noise, and changes in environmental temperature and humidity. The background noise presents time-varying, diverse, and non-stationary characteristics. Traditional fixed parameter filtering methods are difficult to effectively suppress these complex background noises, resulting in low signal-to-noise ratio of the extracted discharge signals and severe distortion of the characteristics. At the same time, traditional time-frequency analysis methods have limited time-frequency resolution, high computational complexity, and are sensitive to noise when processing non-stationary transient discharge signals, making it difficult to accurately extract the complete time-frequency characteristics of weak discharge signals in a strong noise environment. That is, there is a technical problem in the prior art that it is difficult to accurately extract the partial discharge signals of power equipment under complex working conditions. SUMMARY

[0003] Therefore, the present application provides a method for extracting electrical characteristic signals of power equipment under complex working conditions, which can solve the technical problem of difficult accurate extraction of partial discharge signals of power equipment under complex working conditions in the prior art.

[0004] The application is implemented as follows: the application provides a method for extracting electrical characteristic signals of power equipment under complex working conditions, which comprises the following steps: arranging a multi-channel ultrasonic sensor array on the surface of the power equipment to collect ultrasonic signals generated by partial discharge and establishing an original ultrasonic signal sequence; performing adaptive noise cancellation processing on the original ultrasonic signal sequence to obtain a noise-reduced ultrasonic signal sequence; performing five-layer wavelet packet decomposition on the noise-reduced ultrasonic signal sequence to obtain frequency band sub-signals of different frequency bands and constructing an ultra-sparse frequency band energy distribution vector; calling an adaptive time-frequency analysis model to perform time-frequency analysis on each frequency band sub-signal to extract instantaneous frequency parameters, instantaneous amplitude parameters and instantaneous phase parameters and construct a micro-time-frequency feature matrix; performing independent component analysis on the micro-time-frequency feature matrix to obtain independent source signal components and calculating kurtosis values and skewness values; fusing the kurtosis values and the skewness values with the ultra-sparse frequency band energy distribution vector to construct a quasi-stationary comprehensive feature vector; calculating the Euclidean distance between the quasi-stationary comprehensive feature vector and each type of standard discharge characteristic vector in a standard discharge characteristic vector library to determine a partial discharge signal and record discharge type and discharge intensity parameters; selecting a corresponding prediction algorithm according to the discharge mode; and starting a multi-parameter coupling optimization adjustment mode under certain conditions and constructing a final electrical characteristic signal description vector according to the discharge type, the discharge intensity parameters, the quasi-stationary comprehensive feature vector, the instantaneous frequency parameters, the instantaneous amplitude parameters and the instantaneous phase parameters.

[0005] The adaptive noise cancellation processing is specifically to establish a noise estimation model by using a reference noise channel signal and a main signal channel signal, and eliminate background noise components by using a least mean square error iterative algorithm.

[0006] The least mean square error iterative algorithm is specifically to establish a linear relationship model of the background noise components in the reference noise channel signal and the main signal channel signal, and iteratively update filter coefficients by using a gradient descent method so that the sum of squared errors between the filter output and the background noise components in the main signal channel signal is minimized.

[0007] The five-layer wavelet packet decomposition adopts a Daubechies wavelet basis function, and the high-frequency sub-signals and the low-frequency sub-signals obtained after each layer of decomposition are further decomposed, and the Mallat fast algorithm is adopted to realize the decomposition.

[0008] The establishment of the ultra-sparse frequency band energy distribution vector is specifically to record the first frequency band sub-signal after the five-layer wavelet packet decomposition as , calculate the frequency band energy value of the frequency band sub-signal as the sum of the squared amplitudes of all sampling points of the frequency band sub-signal, and record the first frequency band sub-signal after the five-layer wavelet packet decomposition as . ​The band energy values of the layer all band sub-signals are normalized to form an energy distribution sequence, and the band energy value corresponding to the band whose band energy value is greater than 5% of the total energy is selected as an effective energy component, and the band energy values of the remaining bands are set to zero.

[0009] The adaptive time-frequency analysis model adopts an encoder-decoder architecture, the encoder includes four one-dimensional convolution layers, each one-dimensional convolution layer is followed by a batch normalization layer and an activation function layer, the decoder adopts a separable attention mechanism, decomposes the traditional multi-head attention into a spatial attention module and a channel attention module, and adaptively and weightedly fuses through a gating fusion unit.

[0010] The training of the adaptive time-frequency analysis model adopts a robustness enhancement framework based on noise injection training, injects training noise obeying a zero-mean Gaussian distribution on the output feature map of each layer of the encoder, and the strength of the training noise gradually increases from the first layer to the fourth layer.

[0011] The establishment of the micro time-frequency feature matrix is specifically that the intrinsic mode function components are obtained by performing empirical mode decomposition on the denoised ultrasonic signal sequence, the analytic signals are obtained by performing Hilbert transform on each intrinsic mode function component, the instantaneous amplitude parameters and the instantaneous frequency parameters of the analytic signals are calculated, and the matrix is established on a time-frequency plane with time as a row index and frequency as a column index.

[0012] The independent component analysis adopts a FastICA algorithm, separates the signal sources by maximizing the non-Gaussianity of each independent source signal component, calculates the non-Gaussianity by using a negentropy approximation method, and accelerates convergence by using a Newton iteration method in the iteration process.

[0013] The kurtosis value is the ratio of the fourth central moment of the amplitude distribution of the independent source signal component to the square of the second central moment, and the skewness value is the ratio of the third central moment of the amplitude distribution of the independent source signal component to the 1.5th power of the second central moment.

[0014] The establishment of the quasi-stationary comprehensive feature vector is specifically that statistical feature parameters of each independent source signal component in a short time window are extracted, non-zero elements in the super-sparse band energy distribution vector are extracted as frequency domain feature components, and energy barycenter positions of the micro time-frequency feature matrix are calculated according to the time axis direction and the frequency axis direction as time-frequency domain feature components, and the statistical feature parameters, the frequency domain feature components and the time-frequency domain feature components are linearly combined according to normalized weights.

[0015] When the pulse discharge mode is detected and the signal duration length is in the range of ∈[0.1, 2.0] milliseconds, the time-domain peak tracking prediction algorithm is adopted; when the periodic discharge mode is detected, the frequency-domain harmonic analysis prediction algorithm is adopted; and when the random discharge mode is detected, the statistical distribution fitting prediction algorithm is adopted.

[0016] When the high-frequency component proportion of the ultra-sparse frequency band energy distribution vector is greater than 60%, the condition number of the micro-short frequency feature matrix is less than 15, and the variance of the quasi-stationary comprehensive feature vector is less than 0.08, the multi-parameter coupling optimization adjustment mode is started, and the sensor gain coefficient, filter cutoff frequency and Euclidean distance threshold are adjusted by the multivariate coupling algorithm.

[0017] The multivariate coupling algorithm establishes a constraint relationship model among the high-frequency component proportion, the condition number and the variance, converts the constraint optimization problem into an unconstrained optimization problem by using the Lagrange multiplier method, and determines the optimal adjustment amount of each parameter by iteratively solving the extreme point of the Lagrange function.

[0018] The establishment of the final electrical characteristic signal description vector is specifically to encode the discharge type into a type code value, normalize the discharge intensity parameter into an intensity normalized value, extract the effective feature components with a variance greater than 0.01 in the quasi-stationary comprehensive feature vector, extract the statistical mean and statistical variance of the instantaneous frequency parameter, the instantaneous amplitude parameter and the instantaneous phase parameter, and arrange them in order to form the final electrical characteristic signal description vector.

[0019] The number of heads of the separable attention mechanism is determined according to the number of non-zero elements of the ultra-sparse frequency band energy distribution vector, the rank of the micro-short frequency feature matrix and the effective dimension of the quasi-stationary comprehensive feature vector. The number of heads is calculated by taking the weighted average of the three parameters after normalization and multiplying the final number of heads by the reference number of heads 8.

[0020] The application realizes high-precision extraction of partial discharge signals under complex working conditions by constructing a multi-level signal processing architecture and adopting a cooperative processing mechanism of adaptive noise cancellation, wavelet packet multi-scale decomposition, independent component analysis and adaptive time-frequency analysis model. The application establishes a dynamic correlation model of reference noise and noise components in the main signal by using adaptive noise cancellation processing, realizes adaptive tracking and elimination of time-varying background noise through a least mean square error iterative algorithm, and solves the problem that traditional fixed filters cannot cope with complex noise environments. The application constructs an ultra-sparse frequency band energy distribution vector through five-layer wavelet packet decomposition, extracts mutually independent source signal components in combination with independent component analysis, and forms a quasi-stationary comprehensive feature vector by fusing kurtosis value, skewness value and other statistical characteristics, thereby realizing comprehensive representation of multi-dimensional characteristics of partial discharge signals. Meanwhile, an adaptive time-frequency analysis model based on noise injection training is introduced, so that the system can maintain stability and accuracy of feature extraction when facing unknown noise disturbance. In summary, the application solves the technical problem of difficult accurate extraction of partial discharge signals of power equipment under complex working conditions in the background art. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the method of the application.

[0022] Figure 2 is a time-frequency distribution graph of the noise-reduced ultrasonic signal in the embodiment.

[0023] Figure 3 is a distribution graph of kurtosis value and skewness value of the independent source signal component in the embodiment.

[0024] Figure 4 is a visualization graph of the ultra-sparse frequency band energy distribution vector in the embodiment.

[0025] Figure 5 is a time evolution curve graph of the quasi-stationary comprehensive feature vector in the embodiment. DETAILED DESCRIPTION

[0026] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0027] As shown in Figure 1 is a flowchart of an electrical characteristic signal extraction method of power equipment under complex working conditions provided by the application. The method includes the following steps:

[0028] S01, arranging a multi-channel ultrasonic sensor array on the surface of the power equipment, collecting ultrasonic signals generated by partial discharge, and establishing an original ultrasonic signal sequence;

[0029] S02. Adaptive noise cancellation processing is performed on the original ultrasonic signal sequence. A noise estimation model is established using the reference noise channel signal and the main signal channel signal. The background noise component is eliminated by the minimum mean square error iterative algorithm to obtain the noise-reduced ultrasonic signal sequence.

[0030] S03. Perform wavelet packet five-level decomposition on the noise-reduced ultrasonic signal sequence, decompose the noise-reduced ultrasonic signal sequence into 32 frequency band sub-signals of different frequency bands, calculate the frequency band energy value of each frequency band sub-signal and construct the ultra-sparse frequency band energy distribution vector.

[0031] S04. Call the adaptive time-frequency analysis model to perform time-frequency analysis on each frequency band sub-signal, extract the instantaneous frequency parameters, instantaneous amplitude parameters and instantaneous phase parameters of each frequency band sub-signal, and construct the micro-hour frequency feature matrix;

[0032] S05. Perform independent component analysis on the micro-hour frequency characteristic matrix. By iteratively optimizing the separation matrix, the micro-hour frequency characteristic matrix is ​​decomposed into mutually independent source signal components. Calculate the kurtosis and skewness of each independent source signal component.

[0033] S06. The kurtosis and skewness values ​​of each independent source signal component are fused with the ultra-sparse frequency band energy distribution vector to construct a metastable integrated feature vector. The Euclidean distance between the metastable integrated feature vector and the standard discharge feature vectors of each type in the standard discharge feature vector library is calculated.

[0034] S07. When the Euclidean distance is less than the set Euclidean distance threshold, it is determined to be a partial discharge signal, and the discharge type and discharge intensity parameters are recorded. When a pulsed discharge mode is detected and the signal duration is within the range of [0.1, 2.0] milliseconds, a time-domain peak tracking prediction algorithm is used. When a periodic discharge mode is detected, a frequency-domain harmonic analysis prediction algorithm is used. When a random discharge mode is detected, a statistical distribution fitting prediction algorithm is used.

[0035] S08. If the high-frequency component of the ultra-sparse frequency band energy distribution vector accounts for more than 60%, the condition number of the micro-hour frequency characteristic matrix is ​​less than 15, and the variance of the metastable comprehensive characteristic vector is less than 0.08, the multi-parameter coupling optimization adjustment mode is started. The sensor gain coefficient, filter cutoff frequency and Euclidean distance threshold are adjusted through the multi-variable coupling algorithm to achieve coordinated optimization control of the three parameters: high-frequency component proportion, condition number and variance.

[0036] S09. Based on the discharge type, discharge intensity parameter, transient stable comprehensive feature vector, instantaneous frequency parameter, instantaneous amplitude parameter, and instantaneous phase parameter, construct the final electrical characteristic signal description vector.

[0037] The process of establishing the ultra-sparse frequency band energy distribution vector is as follows: the fifth wavelet packet after five-level decomposition is decomposed into the first wavelet packet. layer The frequency band sub-signals are denoted as The band energy value of the frequency band sub-signals is calculated as the sum of the square of the amplitude of each sampling point of the frequency band sub-signals, and the band energy values of all the frequency band sub-signals of the first After normalization processing of the band energy values of all the frequency band sub-signals, an energy distribution sequence is formed, the band energy value corresponding to the frequency band whose band energy value is greater than 5% of the total energy is selected as the effective energy component, the band energy values of the remaining frequency bands are set to zero, and a sparse energy distribution sequence is formed, the energy distribution sequence is the super-sparse frequency band energy distribution vector, the dimension of the super-sparse frequency band energy distribution vector is , the number of non-zero elements in the super-sparse frequency band energy distribution vector is not more than 15% of the total dimension, which reflects the high concentration characteristics of the frequency band energy distribution.

[0038] The establishment process of the micro-time-frequency feature matrix is as follows: the denoised ultrasonic signal sequence is subjected to empirical mode decomposition, and the denoised ultrasonic signal sequence is decomposed into a limited number of intrinsic mode function components, each intrinsic mode function component is subjected to Hilbert transform to obtain an analytical signal, the instantaneous amplitude parameter and the instantaneous frequency parameter of the analytical signal are calculated, and a matrix is established on the time-frequency plane with time as the row index and frequency as the column index, the matrix element is the instantaneous amplitude parameter of the corresponding time and frequency point, due to the transient characteristics of the partial discharge signal, most of the elements in the matrix have values close to zero, only the elements in the discharge time and the corresponding frequency range have significant values, the sparsity of the matrix is more than 85%, so it is called a micro-time-frequency feature matrix, the number of rows of the micro-time-frequency feature matrix is the number of time sampling points, and the number of columns of the micro-time-frequency feature matrix is the number of frequency resolution units.

[0039] The establishment process of the transient stationary comprehensive feature vector is as follows: statistical characteristic parameters of each independent source signal component in a short time window are extracted, the statistical characteristic parameters include six types of parameters including root mean square value, peak factor, waveform factor, margin factor, skewness value and kurtosis value, the non-zero elements in the super-sparse frequency band energy distribution vector are extracted as frequency domain feature components, the energy barycenter position in the time axis direction and the energy barycenter position in the frequency axis direction of the micro-time-frequency feature matrix are calculated as time-frequency domain feature components, the statistical characteristic parameters, the frequency domain feature components and the time-frequency domain feature components are linearly combined according to the normalized weights to form a comprehensive feature vector, the comprehensive feature vector maintains a relatively stable statistical characteristic in a short time window, but slowly changes between different time windows, which reflects the transient stationary characteristics of the signal, so it is called a transient stationary comprehensive feature vector.

[0040] The establishment process of the final electrical characteristic signal description vector is: encoding the discharge type into a type code value, normalizing the discharge intensity parameter into an intensity normalized value, extracting effective feature components with a variance greater than 0.01 in the quasi-stationary comprehensive feature vector, extracting the statistical mean and statistical variance of the instantaneous frequency parameter as frequency statistical features, extracting the statistical mean and statistical variance of the instantaneous amplitude parameter as amplitude statistical features, extracting the statistical mean and statistical variance of the instantaneous phase parameter as phase statistical features, arranging the type code value, the intensity normalized value, the effective feature components, the frequency statistical features, the amplitude statistical features and the phase statistical features in order to form the final electrical characteristic signal description vector, and the final electrical characteristic signal description vector is used to represent the complete electrical characteristic information of partial discharge and is used for subsequent fault diagnosis and state evaluation.

[0041] The structure of the adaptive time-frequency analysis model is: the adaptive time-frequency analysis model adopts an encoder-decoder architecture, the encoder includes four one-dimensional convolution layers, each one-dimensional convolution layer is followed by a batch normalization layer and an activation function layer, the convolution kernel size of the one-dimensional convolution layer is 64, 128, 256 and 512 respectively, and the step length of the one-dimensional convolution layer is 2, which is used to extract multi-scale time domain features of the denoising ultrasonic signal sequence; the decoder adopts a separable attention mechanism, which decomposes the traditional multi-head attention into a spatial attention module and a channel attention module, the spatial attention module calculates the dependency between different positions in the time dimension, the channel attention module calculates the dependency between different feature channels, and the output of the spatial attention module and the output of the channel attention module are adaptively weighted and fused through a gating fusion unit; the number of heads of the separable attention mechanism in the adaptive time-frequency analysis model is determined according to three parameters: the number of non-zero elements of the ultra-sparse frequency band energy distribution vector, the rank of the micro-time-frequency feature matrix, and the effective dimension of the quasi-stationary comprehensive feature vector, the number of heads is calculated by normalizing the three parameters, then taking a weighted average, and then multiplying the reference number of heads 8 to obtain the final number of heads; the output layer of the adaptive time-frequency analysis model is a full connection layer, and the output dimension of the output layer is the size of the micro-time-frequency feature matrix, and the time-frequency distribution is reconstructed through inverse transformation.

[0042] The step of establishing the training data set of the adaptive time-frequency analysis model comprises: building a partial discharge simulation device in a laboratory environment, simulating three typical discharge types of internal discharge, surface discharge and corona discharge respectively, collecting ultrasonic signals under different discharge voltage levels and different discharge gap distances, and collecting not less than 5000 groups of effective discharge signal samples for each discharge type; manually labeling the collected ultrasonic signals, and the labeling content comprises discharge type, discharge starting time, discharge duration and discharge intensity parameter; dividing the labeled ultrasonic signals into a training set and a validation set according to an 8:2 ratio; performing data enhancement processing on the ultrasonic signals in the training set, wherein the data enhancement processing comprises adding Gaussian white noise of different intensities, adding power frequency interference signals, time shifting and amplitude scaling, so that the number of training set samples is expanded to 5 times the number of original samples.

[0043] The step of training the adaptive time-frequency analysis model comprises: adopting a weighted combination of mean square error loss function and cross-entropy loss function as a total loss function, the mean square error loss function is used to constrain the waveform error of the reconstructed signal and the original ultrasonic signal sequence, and the cross-entropy loss function is used to constrain the accuracy of discharge type classification; adopting AdamW optimizer for parameter updating, setting the initial learning rate to 0.001, and attenuating the learning rate to 0.5 times of the original after every 20 rounds of training; adopting a robustness enhancement framework based on noise injection training in the training process, injecting training noise conforming to zero-mean Gaussian distribution on each layer output feature map of the encoder, the intensity of the training noise gradually increases from the first layer to the fourth layer, and the standard deviation of the training noise is set to 0.05 times, 0.10 times, 0.15 times and 0.20 times of the standard deviation of the feature map respectively, by forcing the adaptive time-frequency analysis model to still be able to accurately extract time-frequency features under training noise disturbance, the resistance of the adaptive time-frequency analysis model to inherent noise and external interference in the input signal is improved; the performance of the adaptive time-frequency analysis model is evaluated on the validation set every 5 rounds in the training process, and the training is stopped in advance when the validation set loss does not decrease for 10 consecutive rounds.

[0044] The robustness enhancement framework based on noise injection training constructs an active robust learning mechanism by systematically injecting controllable intensity of training noise disturbance at different levels of the adaptive time-frequency analysis model, essentially changing the learning paradigm of the adaptive time-frequency analysis model, so that the adaptive time-frequency analysis model not only learns the direct mapping relationship from clean signals to target outputs, but more importantly learns the internal mechanism of keeping feature extraction stable in the training noise environment. The injected training noise forces the adaptive time-frequency analysis model to find a robust decision boundary in the presence of disturbance in the feature space during the forward propagation process. The learning process of the training noise is essentially a regularization method that effectively suppresses the overfitting of the adaptive time-frequency analysis model to the noise patterns in the training data, so that the feature representation learned by the adaptive time-frequency analysis model has stronger generalization ability. When the adaptive time-frequency analysis model is deployed in the actual working environment and faces unknown noise types, since the noise-robust feature extraction mechanism has been established during the training stage, the adaptive time-frequency analysis model can automatically suppress the interference of noise on feature extraction and maintain the stability of the output. In addition, training noise injection training also enhances the activation diversity of neurons in the adaptive time-frequency analysis model, avoids over-sensitivity of neurons to input patterns, and makes the adaptive time-frequency analysis model exhibit smoother response characteristics as a whole. The mechanism is particularly crucial for the partial discharge signal extraction scene of power equipment, because the background noise components in the actual working condition have time-varying and diversity characteristics, and traditional training methods are difficult to cover all background noise component conditions. Training noise injection training simulates the influence of training noise on feature propagation, so that the adaptive time-frequency analysis model obtains the internal ability to cope with unknown noise disturbance, thereby significantly improving the reliability and accuracy of the entire signal extraction method in harsh electromagnetic environments and strong background noise conditions.

[0045] The calculation efficiency optimization algorithm based on separable attention fundamentally reduces the calculation complexity by decomposing the global attention calculation in the traditional multi-head attention mechanism into two independent attention calculation processes in the spatial dimension and the channel dimension. The traditional multi-head attention mechanism needs to calculate the similarity matrix between the query vector and all key vectors, and the calculation complexity is proportional to the square of the length of the input sequence, which leads to a sharp increase in the consumption of computing resources when processing long-time sequence of denoising ultrasonic signals. The separable attention algorithm first calculates the dependency between different time positions in the spatial dimension, which only needs to calculate the attention weight in the time axis direction, and the complexity of this step is linearly related to the sequence length. Then, the dependency between different feature channels is calculated in the channel dimension, and the calculation amount of this step is only related to the number of feature channels. The attention weights of the two dimensions are adaptively fused through a gating fusion unit. The overall calculation complexity is reduced from the square level of the traditional multi-head attention mechanism to the linear level. The decomposition not only greatly reduces the calculation amount, but more importantly, it retains the modeling ability of the attention mechanism for long-range dependencies. Because the spatial attention module captures the correlation between features at different times in the time sequence, and the channel attention module captures the coupling relationship between different frequency components, the synergistic effect of the two achieves comprehensive modeling of the time-frequency features of the denoising ultrasonic signal sequence. In practical applications, the optimization algorithm enables the adaptive time-frequency analysis model to run in real time on an embedded hardware platform, meeting the real-time requirements of power equipment online monitoring. At the same time, the reduction in calculation amount reduces the energy consumption of the adaptive time-frequency analysis model, prolonging the battery life of the sensor node, which has important practical value for distributed power equipment monitoring systems. In addition, the improvement of calculation efficiency also enables the adaptive time-frequency analysis model to process longer-time window denoising ultrasonic signal sequences and capture more complete discharge process evolution information, further improving the accuracy and integrity of fault feature extraction.

[0046] The least mean square error iterative algorithm in the adaptive noise cancellation process establishes a linear relationship model between the background noise components in the reference noise channel signal and the main signal channel signal, iteratively updates the filter coefficients using the gradient descent method, and minimizes the error sum of squares between the filter output and the background noise components in the main signal channel signal, thereby achieving adaptive estimation and elimination of the background noise components.

[0047] The wavelet packet five-layer decomposition uses the Daubechies wavelet basis function, and the decomposition layer is set to 5 layers. The high-frequency sub-signals and low-frequency sub-signals obtained after each layer of decomposition are further decomposed, and finally 32 frequency band sub-signals of different frequency bands are obtained. The wavelet packet five-layer decomposition process is realized by using the Mallat fast algorithm.

[0048] The independent component analysis adopts a FastICA algorithm, signal source separation is realized by maximizing the non-Gaussianity of each independent source signal component, the non-Gaussianity measure is calculated by using a negative entropy approximation method, and the Newton iteration method is used in the iteration process to accelerate convergence, and the convergence criterion is that the change of the separation matrix in the adjacent two iterations is less than 0.0001.

[0049] The kurtosis value is the ratio of the fourth-order central moment to the square of the second-order central moment of the amplitude distribution of the independent source signal component, and reflects the sharpness of the amplitude distribution of the independent source signal component, and the kurtosis value greater than 3 indicates that the independent source signal component has a sharp distribution characteristic, and the kurtosis value less than 3 indicates that the independent source signal component has a flat distribution characteristic.

[0050] The skewness value is the ratio of the third-order central moment to the square of the second-order central moment of the amplitude distribution of the independent source signal component, and reflects the symmetry of the amplitude distribution of the independent source signal component, and the skewness value greater than 0 indicates that the independent source signal component distribution is right-skewed, and the skewness value less than 0 indicates that the independent source signal component distribution is left-skewed.

[0051] The time-domain peak tracking prediction algorithm identifies the local maximum value point in the noise-reduced ultrasonic signal sequence, calculates the time interval and amplitude ratio between adjacent local maximum value points, and predicts the time and amplitude of the next peak value using a linear extrapolation method.

[0052] The frequency-domain harmonic analysis prediction algorithm performs fast Fourier transform on the noise-reduced ultrasonic signal sequence, identifies the main harmonic component and its frequency position in the frequency spectrum, and predicts the signal spectrum distribution at future time according to the phase evolution law of the main harmonic component.

[0053] The statistical distribution fitting prediction algorithm calculates the probability distribution characteristics of the noise-reduced ultrasonic signal sequence amplitude, fits the parameters of the Weibull distribution or lognormal distribution using the maximum likelihood estimation method, and predicts the probability distribution range of the noise-reduced ultrasonic signal sequence amplitude at future time according to the parameters of the fitted distribution.

[0054] The multivariate coupling algorithm establishes a constraint relationship model among the high-frequency component proportion, the condition number and the variance, converts the constraint optimization problem into an unconstrained optimization problem by using the Lagrange multiplier method, and determines the optimal adjustment amount of each parameter by iteratively solving the extreme point of the Lagrange function.

[0055] The sensor gain coefficient is the proportional coefficient between the output voltage of the ultrasonic sensor array and the input sound pressure, and the unit of the sensor gain coefficient is millivolt per Pascal. Adjusting the sensor gain coefficient changes the sensitivity of the ultrasonic sensor array, and reducing the sensor gain coefficient when the high-frequency component proportion is too large can avoid saturation of the original ultrasonic signal sequence.

[0056] The filter cutoff frequency is the boundary frequency of the frequency range allowed to pass through the band-pass filter, the unit of the filter cutoff frequency is kilohertz, the filter cutoff frequency is dynamically adjusted according to the condition number of the micro-time-frequency feature matrix, and the range of the passband is expanded to retain more effective information when the condition number is small.

[0057] The high-frequency component proportion is the ratio of the sum of the frequency band energy values of the frequency band sub-signals with a frequency greater than 100 kilohertz to the total sum of the frequency band energy values of all frequency band sub-signals in the super-sparse frequency band energy distribution vector, and the high-frequency component proportion reflects the richness of the high-frequency component of the denoised ultrasonic signal sequence.

[0058] The condition number is the ratio of the maximum singular value to the minimum singular value of the micro-time-frequency feature matrix, and the condition number reflects the numerical stability and independence of the micro-time-frequency feature matrix, and the smaller the condition number is, the weaker the linear correlation between the rows or columns of the micro-time-frequency feature matrix is.

[0059] The variance is the variance of each element of the quasi-stationary comprehensive feature vector, and the variance reflects the fluctuation degree of the quasi-stationary comprehensive feature vector in a short time window, and the smaller the variance is, the more stable the quasi-stationary comprehensive feature vector is.

[0060] The pulse discharge mode is a discharge mode in which the discharge signal presents single or a small number of pulse characteristics, and the pulse discharge mode is characterized by a sharp peak in the instantaneous amplitude parameter and a short duration.

[0061] The periodic discharge mode is a discharge mode in which the discharge signal presents regular repetition characteristics, and the periodic discharge mode is characterized by an obvious harmonic component distribution in the frequency spectrum of the instantaneous frequency parameter.

[0062] The random discharge mode is a discharge mode in which the discharge signal presents irregular random characteristics, and the random discharge mode is characterized by no obvious regularity in the instantaneous amplitude parameter and the instantaneous frequency parameter.

[0063] The type coding value is a coding value that converts the discharge type into a numerical form, the internal discharge coding is 1, the surface discharge coding is 2, and the corona discharge coding is 3.

[0064] The effective feature component is a feature component in the quasi-stationary comprehensive feature vector that has a significant contribution to fault diagnosis, and the selection criterion of the effective feature component is that the variance of the feature component is greater than 0.01.

[0065] The frequency statistical feature includes a statistical mean and a statistical variance of the instantaneous frequency parameter, the statistical mean reflects the central tendency of the instantaneous frequency parameter, and the statistical variance reflects the dispersion degree of the instantaneous frequency parameter.

[0066] The amplitude statistical feature includes a statistical mean and a statistical variance of the instantaneous amplitude parameter, the statistical mean reflecting a central tendency of the instantaneous amplitude parameter, and the statistical variance reflecting a dispersion degree of the instantaneous amplitude parameter.

[0067] The phase statistical feature includes a statistical mean and a statistical variance of the instantaneous phase parameter, the statistical mean reflecting a central tendency of the instantaneous phase parameter, and the statistical variance reflecting a dispersion degree of the instantaneous phase parameter.

[0068] As an optional implementation, the application further provides a computer- implemented manner to form an electrical feature signal extraction system of a power equipment, the computer is provided with a readable storage medium, the readable storage medium is stored with program instructions, and the program instructions execute the above method when running in the computer.

[0069] The specific implementation of the above steps is described in detail below.

[0070] The specific implementation of step S01 is to arrange a plurality of ultrasonic sensors at a predetermined interval on the surface of a key part of the power equipment to form an array structure, the sensor array adopts a resonant ultrasonic sensor made of piezoelectric ceramic material, the center frequency is selected as 150 kHz to cover the main frequency band of the ultrasonic wave generated by partial discharge, the interval between the sensors is set as 30-50 mm to ensure that the spatial sampling rate meets the positioning accuracy requirement, each sensor is connected to a multi-channel data acquisition system through a low-noise preamplifier, the sampling frequency of the acquisition system is set as 1 MHz to meet the Nyquist sampling theorem requirement, the acquisition duration is set as 10-100 ms according to the discharge type, the time alignment of the signals of each channel is ensured through a synchronous trigger mechanism, and the acquired ultrasonic signals of each channel are arranged in time sequence to form an original ultrasonic signal sequence, which contains the partial discharge signal and various background noise components.

[0071] The specific implementation of step S02 is to select a sensor channel far from the discharge source position as a reference noise channel, which mainly collects background noise components and weak partial discharge signal components, and to select a sensor channel close to the discharge source position as a main signal channel, which contains both partial discharge signals and background noise components. An adaptive filter structure is established, the input of the filter is the reference noise channel signal, the filter estimates the background noise components in the main signal channel by linear combination of the reference noise channel signal through adjustable weight coefficients, the least mean square algorithm is used to iteratively update the filter weight coefficients, the algorithm adjusts the weight coefficients by calculating the correlation between the error signal and the reference noise channel signal based on the gradient descent principle, the iteration step size parameter is set to 0.01 to balance the convergence speed and stability, and the filter order is set to 32 to fully model the spectral characteristics of the background noise components. The noise estimation value output by the filter is subtracted from the main signal channel signal to obtain a denoised ultrasonic signal sequence, in which the background noise components are effectively suppressed and the partial discharge signals are retained.

[0072] The specific implementation of step S03 is to perform wavelet packet decomposition on the denoised ultrasonic signal sequence using the Daubechies wavelet basis function, which has good time-frequency localization characteristics and orthogonality. The decomposition process is realized by using the Mallat fast algorithm, which recursively decomposes the signal by cascading low-pass and high-pass filters. The first layer decomposition divides the denoised ultrasonic signal sequence into one low-frequency sub-signal and one high-frequency sub-signal, the second layer decomposition further decomposes the two sub-signals of the first layer into four sub-signals, and so on until the fifth layer decomposition obtains 32 frequency band sub-signals of different frequency bands. The frequency band width of each frequency band sub-signal is about 15.625 kHz. The energy value of each frequency band sub-signal is calculated, which is equal to the sum of the squares of the amplitudes of the sampling points. The 32 frequency band energy values are normalized by dividing the total energy to obtain an energy distribution sequence. The energy values of the frequency bands with energy values greater than 5% of the total energy are selected as effective energy components and retained, and the energy values of the remaining frequency bands are set to zero to realize sparse processing. The 32-dimensional sparse energy distribution sequence formed is the ultra-sparse frequency band energy distribution vector, in which the number of non-zero elements is usually not more than 5, which reflects the high concentration of partial discharge signal energy in the frequency domain.

[0073] The specific implementation of step S04 is to input the noise reduction ultrasonic signal sequence into an adaptive time-frequency analysis model, which adopts an encoder-decoder architecture. The encoder includes 4 one-dimensional convolution layers, and the convolution kernel sizes are 64, 128, 256 and 512 respectively, and the step is 2. After each convolution, a batch normalization layer is connected to accelerate the training convergence and improve the model generalization ability. The activation function adopts a rectified linear unit to introduce a nonlinear transformation capability. The encoder extracts the multi-scale time domain features of the noise reduction ultrasonic signal sequence layer by layer and compresses the feature dimension. The decoder adopts a separable attention mechanism. The mechanism decomposes the traditional multi-head attention into two independent modules, a spatial attention module and a channel attention module. The spatial attention module calculates the dependency relationship weight between different position features in the time dimension. The channel attention module calculates the dependency relationship weight between different feature channels. The outputs of the two modules are adaptively weighted and fused through a gating fusion unit. The gating unit dynamically adjusts the contribution proportion of the two attention branches according to the feature statistical characteristics. The output layer of the model is a full connection layer, which maps the fused features to the time-frequency plane to obtain a micro time-frequency feature matrix. The row index of the matrix corresponds to the time sampling point, the column index corresponds to the frequency resolution unit, and the matrix element value represents the instantaneous amplitude parameter at the corresponding time and frequency point. Due to the transient characteristics of the partial discharge signal, the sparsity of the matrix is more than 85%. The instantaneous frequency parameter, the instantaneous amplitude parameter and the instantaneous phase parameter of each frequency band sub-signal are extracted from the micro time-frequency feature matrix. The instantaneous frequency parameter is obtained by calculating the derivative of the phase with respect to time. The instantaneous amplitude parameter is directly read from the matrix element value. The instantaneous phase parameter is obtained by calculating the phase angle of the analytic signal through Hilbert transform.

[0074] The specific implementation of step S05 is to input the micro-time-frequency feature matrix as an observation signal into an independent component analysis algorithm, which is implemented based on a fast independent component analysis method. The method separates the signal sources by maximizing the non-Gaussianity of each independent source signal component. The non-Gaussianity is calculated using a negentropy approximation method. The negentropy is defined as the difference between the signal entropy and the entropy of a homoscedastic Gaussian distribution. The iterative process uses the Newton iteration method to accelerate the convergence speed. The initial separation matrix is the unit matrix. In each iteration, the non-Gaussianity measure and its gradient and Hessian matrix corresponding to the current separation vector are calculated. The separation vector is adjusted according to the Newton method update formula. The separation vector is orthogonalized to ensure the statistical independence of each independent component. The iteration termination condition is that the change in the separation matrix between the last two iterations is less than 0.0001. After the iteration converges, the separation matrix obtained separates the micro-time-frequency feature matrix into multiple independent source signal components. The kurtosis value and skewness value of each independent source signal component are calculated. The kurtosis value is equal to the fourth-order central moment of the amplitude distribution of the independent source signal component divided by the square of the second-order central moment. This statistical quantity reflects the sharpness of the distribution. A kurtosis value greater than 3 indicates a sharp distribution characteristic, and a kurtosis value less than 3 indicates a flat distribution characteristic. The skewness value is equal to the third-order central moment of the amplitude distribution of the independent source signal component divided by the 1.5th power of the second-order central moment. This statistical quantity reflects the symmetry of the distribution. A skewness value greater than 0 indicates a right-skewed distribution, and a skewness value less than 0 indicates a left-skewed distribution.

[0075] The specific implementation of step S06 is to extract the statistical feature parameters of each independent source signal component in a short-time window. The short-time window length is set to 5ms to ensure the transient stationarity of the features. The statistical feature parameters include six types of parameters: root mean square value, peak factor, waveform factor, margin factor, skewness value, and kurtosis value. The non-zero elements from the super-sparse frequency band energy distribution vector are extracted as frequency domain feature components. The energy barycenter positions are calculated in the time axis direction and the frequency axis direction from the micro-time-frequency feature matrix as time-frequency domain feature components. The energy barycenter positions are calculated by the weighted average method. The statistical feature parameters, frequency domain feature components, and time-frequency domain feature components are linearly combined according to the normalized weights. The weight coefficients are determined according to the contribution of each feature to fault diagnosis. The sum of the weight coefficients is normalized to 1. The resulting comprehensive feature vector maintains relatively stable statistical properties in a short time window but slowly changes between different time windows, reflecting the transient stationary characteristics of the signal. This is called a transient stationary comprehensive feature vector. The Euclidean distance between the transient stationary comprehensive feature vector and each type of standard discharge feature vector in the standard discharge feature vector library is calculated. The Euclidean distance is equal to the square root of the sum of the squares of the differences between the corresponding elements of the two vectors. The smaller the Euclidean distance, the more similar the two vectors.

[0076] The specific implementation of step S07 is to set the Euclidean distance threshold value to 0.3, determine the current signal as a partial discharge signal when the calculated Euclidean distance is less than the threshold value, determine the discharge type according to which type of standard discharge characteristic vector has the minimum Euclidean distance, the discharge type including three types of internal discharge, surface discharge and corona discharge, calculate the discharge intensity parameter according to the peak factor and the margin factor in the temporary stationary comprehensive characteristic vector, the discharge intensity parameter reflecting the energy size of the discharge, judge the discharge mode according to the time domain waveform characteristics of the discharge signal, when the pulse type discharge mode is detected and the signal duration length is located in the range of 0.1ms to 2.0ms, the time domain peak tracking prediction algorithm is used, the algorithm identifies the local maximum points in the denoised ultrasonic signal sequence, calculates the time interval and amplitude ratio relationship between adjacent local maximum points, and uses the linear extrapolation method to predict the time and amplitude of the next peak, when the periodic discharge mode is detected, the frequency domain harmonic analysis prediction algorithm is used, the algorithm performs fast Fourier transform on the denoised ultrasonic signal sequence, identifies the main harmonic components and their frequency positions in the frequency spectrum, and predicts the signal frequency spectrum distribution at future time according to the phase evolution law of the main harmonic components, when the random discharge mode is detected, the statistical distribution fitting prediction algorithm is used, the algorithm statistically analyzes the probability distribution characteristics of the amplitude of the denoised ultrasonic signal sequence, fits the parameters of Weibull distribution or lognormal distribution by using the maximum likelihood estimation method, and predicts the probability distribution range of the amplitude of the denoised ultrasonic signal sequence at future time according to the parameters of the fitted distribution.

[0077] The specific implementation of step S08 is to calculate the ratio of the sum of the band energy values of the frequency band sub-signals with a frequency greater than 100 kHz in the super-sparse frequency band energy distribution vector to the total sum of the band energy values of all frequency band sub-signals to obtain the high-frequency component proportion, calculate the ratio of the maximum singular value to the minimum singular value of the micro-time-frequency feature matrix to obtain the condition number, the singular values are obtained by singular value decomposition method, calculate the variance of each element of the quasi-stationary comprehensive feature vector, the variance is equal to the square sum of the difference between each element and the mean value divided by the number of elements, when the three conditions of high-frequency component proportion greater than 60%, condition number less than 15, and variance less than 0.08 are met at the same time, start the multi-parameter coupling optimization adjustment mode, establish the constraint relationship model among the three parameters of high-frequency component proportion, condition number and variance, this model uses the Lagrange multiplier method to convert the constrained optimization problem into an unconstrained optimization problem, constructs the Lagrange function, which contains the objective function term and the constraint condition term, determine the optimal adjustment amount of each parameter by iteratively solving the equation set whose partial derivative of the Lagrange function with respect to each parameter is zero, adjust the sensor gain coefficient to change the sensitivity of the ultrasonic sensor array, the unit of the sensor gain coefficient is millivolt per pascal, the reference value range is 50mV / Pa to 200mV / Pa, reduce the sensor gain coefficient when the high-frequency component proportion is too large to avoid the saturation of the original ultrasonic signal sequence, adjust the filter cutoff frequency to change the frequency range allowed to pass through the band-pass filter, the unit of the filter cutoff frequency is kilohertz, the reference value range is 80kHz to 250kHz, expand the passband range to retain more effective information when the condition number is small, adjust the Euclidean distance threshold to change the sensitivity of partial discharge signal determination, the reference value range of the Euclidean distance threshold is 0.2 to 0.5, realize the coordinated optimization control of the three parameters of high-frequency component proportion, condition number and variance through the multivariate coupling algorithm.

[0078] The specific implementation of step S09 is to encode the discharge type as a type code value, encode internal discharge as 1, encode surface discharge as 2, and encode corona discharge as 3; normalize the discharge intensity parameter by dividing it by a set maximum discharge intensity reference value to obtain an intensity normalized value, the maximum discharge intensity reference value being determined according to the rated voltage level of the power equipment; select, from the temporary stationary comprehensive feature vector, a feature component whose variance is greater than 0.01 as an effective feature component, which removes redundant features that have less contribution to fault diagnosis; calculate the statistical mean and statistical variance of the instantaneous frequency parameter as frequency statistical features, the statistical mean reflecting the central tendency of the instantaneous frequency parameter and the statistical variance reflecting the dispersion degree of the instantaneous frequency parameter; calculate the statistical mean and statistical variance of the instantaneous amplitude parameter as amplitude statistical features; calculate the statistical mean and statistical variance of the instantaneous phase parameter as phase statistical features; arrange the type code value, the intensity normalized value, the effective feature component, the frequency statistical features, the amplitude statistical features, and the phase statistical features in order to form a final electrical feature signal description vector, which is used to represent the complete electrical feature information of the partial discharge and can be used for subsequent fault diagnosis and state evaluation, the vector dimension being dynamically determined according to the number of effective feature components, usually being 20 to 40 dimensions.

[0079] It should be noted that the key technical ideas of the present application include two core technologies of adaptive time-frequency analysis based on separable attention mechanism and robustness enhancement based on noise injection training. The separable attention mechanism breaks through the bottleneck of the calculation complexity increasing with the square of the sequence length by decomposing the global calculation of the traditional multi-head attention into independent calculation processes of spatial dimension and channel dimension, so that the model can realize linear level of calculation complexity while maintaining the ability to model long-range dependencies. This decomposition not only significantly reduces the consumption of computing resources and energy, but also enables the model to run in real time on embedded hardware platforms and process longer time window signals, thereby capturing more complete discharge process evolution information. The robustness enhancement framework based on noise injection training builds an active robust learning mechanism by systematically injecting controllable intensity of training noise disturbance during model training. This mechanism forces the model to learn the internal ability to remain stable in a noisy environment during feature propagation. Essentially, it realizes a regularization method that effectively suppresses the overfitting of the model to noise patterns in the training data, making the feature representation learned by the model have stronger generalization ability and noise resistance. The synergy of the two technical ideas lies in the fact that the separable attention mechanism provides an efficient feature extraction architecture foundation, enabling the model to quickly process long-time signals and extract rich time-frequency features, while the noise injection training further enhances the model's ability to adapt to time-varying noise and unknown interference in complex working conditions. The combined effect of the two makes the entire signal extraction method significantly improve in terms of computational efficiency, feature representation ability, and noise robustness, making it particularly suitable for online monitoring scenarios of power equipment in harsh electromagnetic environments and strong background noise conditions.

[0080] Specifically, the principle of this invention is as follows: This invention solves the technical problem of accurately extracting partial discharge signals from power equipment under complex operating conditions. Its fundamental principle lies in establishing a collaborative processing mechanism of adaptive noise suppression and multi-domain feature fusion. First, adaptive noise cancellation processing establishes a linear relationship model between the reference noise channel and the main signal channel, and dynamically updates the filter coefficients using a minimum mean square error iterative algorithm. This allows the noise estimation to adaptively adjust to the time-varying characteristics of the actual noise, more effectively eliminating background noise under complex operating conditions compared to traditional fixed-parameter filters. Second, wavelet packet five-level decomposition decomposes the signal into 32 sub-signals of different frequency bands, constructing an ultra-sparse frequency band energy distribution vector. Combined with independent component analysis, source separation of aliased signals is performed, comprehensively characterizing the discharge signal features from three dimensions: time domain, frequency domain, and time-frequency domain. This multi-domain feature fusion method can fully explore the inherent structural information of the signal. Finally, the adaptive time-frequency analytical model adopts an encoder-decoder architecture combined with a separable attention mechanism. By injecting noise perturbations of different intensities during the training phase, the model is forced to learn the intrinsic mechanism of maintaining feature extraction stability in noisy environments. This noise-robust training strategy enables the model to generalize in the face of unknown noise perturbations. When deployed in a real working environment and facing complex background noise, the model can automatically suppress noise interference and maintain output stability. Therefore, the technical solution of this invention has inherent logical rationality and technical feasibility.

[0081] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0082] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0083] The specific implementation of step S02 involves selecting a sensor channel far from the power source as the reference noise channel and a sensor channel close to the power source as the main signal channel. An adaptive filter structure is established, and the filter weight coefficients are updated using a minimum mean square error iterative algorithm. The noise estimate of the filter output is expressed as follows:

[0084] ;

[0085] In the formula, for The estimated background noise component at time 1, in millivolts; For the first Each filter weight coefficient is dimensionless. For the reference noise channel signal in The sampled value at any given time, in millivolts; This is the filter order, with a value of 32. This is the index of the filter weight coefficients, with values ​​ranging from 0 to... ; The current time is expressed in seconds.

[0086] The formula for updating the filter weight coefficients is expressed as follows:

[0087] ;

[0088] In the formula, For the first Each filter weight coefficient The update value at any given time is dimensionless. This is the iteration step size parameter, with a value of 0.01, and is dimensionless. for The error signal at time t is represented by the main signal channel signal. Subtract noise estimate The unit is millivolt.

[0089] The formula for calculating the noise-reduced ultrasonic signal sequence is as follows:

[0090] ;

[0091] In the formula, for The noise-reduced ultrasonic signal sequence value at any given time, in millivolts; for The signal value of the main signal channel at any given time, in millivolts.

[0092] The specific implementation of step S03 is to perform a five-level wavelet packet decomposition on the noise-reduced ultrasonic signal sequence. Layer The formula for calculating the frequency band energy value of a sub-band signal is as follows:

[0093] ;

[0094] In the formula, For the first Layer The frequency band energy value of each frequency band sub-signal, in millivolt squared; For the first Layer The frequency band sub-signal in the first The amplitude of each sampling point, in millivolts; This represents the total number of sampling points for the frequency band sub-signal; This represents the wavelet packet decomposition level, with a value of 5. This is the index of the frequency band sub-signal, with a value range of 0 to... ; This is the sampling point index, with a value range of 1 to... .

[0095] The formula for calculating the normalized energy distribution sequence is as follows:

[0096] ;

[0097] In the formula, For the first Layer The normalized frequency band energy value of each frequency band sub-signal, dimensionless; For the first The total number of frequency band sub-signals of the layer, when The time is 32.

[0098] The formula for constructing the energy distribution vector of the ultra-sparse frequency band is as follows:

[0099] ;

[0100] In the formula, The first element in the ultra-sparse frequency band energy distribution vector One element, dimensionless; threshold 0.05 represents 5% of the total energy.

[0101] The specific implementation of step S04 involves inputting the noise-reduced ultrasonic signal sequence into an adaptive time-frequency analytical model. The model extracts the microsecond hourly frequency feature matrix, and the calculation formula for the instantaneous frequency parameter is expressed as follows:

[0102] ;

[0103] In the formula, for The instantaneous frequency parameter at a given time, in Hertz; for The instantaneous phase parameter at a given moment, in radians; This is the first derivative of the instantaneous phase parameter with respect to time, expressed in radians per second.

[0104] The formula for constructing the micro-hour frequency feature matrix is ​​as follows:

[0105] ;

[0106] In the formula, For micro-hour frequency feature matrix in row index Column index The element value at position, in millivolts; For a moment and frequency The corresponding instantaneous amplitude parameter, in millivolts; This is the index of the time sampling point, with a value range from 1 to... ; This is the frequency resolution unit index, with a value range of 1 to... ; This represents the total number of time sampling points; This represents the total number of frequency resolution units. For the first The time corresponding to each time sampling point is in seconds; For the first The frequency value corresponding to each frequency resolution unit is expressed in Hertz.

[0107] The specific implementation of step S05 involves using the micro-hour frequency characteristic matrix as the input to the fast independent component analysis algorithm, and the calculation formulas for each order of central moments are expressed as follows:

[0108] ;

[0109] In the formula, for The first central moment, when The unit is millivolt squared, when The unit is millivolts per cubic meter. The unit of time is millivolts to the power of 4; The first independent source signal component The amplitude of each sampling point is expressed in millivolts. This represents the average amplitude of the independent source signal components, in millivolts. The total number of sampling points for independent source signal components; The order of the central moment is 2, 3, or 4.

[0110] The formula for calculating kurtosis is as follows:

[0111] ;

[0112] In the formula, Kurtosis is a dimensionless value. The fourth central moment of the amplitude distribution of independent source signal components, in millivolts to the power of 4; It represents the second-order central moment of the amplitude distribution of independent source signal components, in millivolt squared units.

[0113] The formula for calculating the skewness value is expressed as follows:

[0114] ;

[0115] In the formula, This is the skewness value, which is dimensionless. It represents the third-order central moment of the amplitude distribution of independent source signal components, in millivolt cubic units.

[0116] The specific implementation of step S06 involves extracting the statistical characteristic parameters of each independent source signal component within a short time window. The short time window length is set to 5 milliseconds, and the calculation formula for the statistical characteristic parameters is as follows:

[0117] The formula for calculating the root mean square value is:

[0118] ;

[0119] The formula for calculating the peak factor is:

[0120] ;

[0121] The formula for calculating the waveform factor is:

[0122] ;

[0123] The formula for calculating the margin factor is:

[0124] ;

[0125] In the formula, This is the root mean square value, in millivolts. The peak factor is dimensionless. The waveform factor is dimensionless. The margin factor is dimensionless. For the first time window The amplitude of the independent source signal components at each sampling point, in millivolts; This represents the total number of sampling points within the short time window; This represents the maximum absolute value of the amplitude of the independent source signal component within the short time window, in millivolts. This is the index of the sampling points within the short time window, with a value ranging from 1 to... .

[0126] The formula for calculating the location of the energy centroid in the time-frequency domain is as follows:

[0127] ;

[0128] ;

[0129] In the formula, This is the normalized value of the energy centroid position along the time axis, and is dimensionless. This is the normalized value of the energy centroid position along the frequency axis, and is dimensionless. This is the time normalization reference value, which is the total duration of the signal, in seconds. This is the frequency normalization reference value, which is the maximum frequency value, and the unit is Hertz.

[0130] The formula for constructing the metastable comprehensive eigenvector is expressed as follows:

[0131] ;

[0132] wherein, is a temporary stationary comprehensive feature vector, with dimension ; is a statistical feature parameter vector, including 6 elements of root mean square value, peak factor, waveform factor, margin factor, skewness value and kurtosis value; is a non-zero element vector in the super-sparse frequency band energy distribution vector; is a normalized weight coefficient, dimensionless, satisfying , and the empirical values are 0.3, 0.4, 0.15 and 0.15, respectively; is the dimension of the temporary stationary comprehensive feature vector; the superscript represents the vector transposition operation.

[0133] The calculation formula of the Euclidean distance is as follows:

[0134] ;

[0135] wherein, is the Euclidean distance, dimensionless; is the th element of the temporary stationary comprehensive feature vector, dimensionless; is the th element of the standard discharge feature vector, dimensionless; is a feature vector element index, with a value range of 1 to .

[0136] The specific implementation of step S07 is the same as the foregoing, and will not be described in detail here.

[0137] The specific implementation of step S08 is to calculate the high-frequency component proportion, and the calculation formula is as follows:

[0138] ;

[0139] wherein, is the high-frequency component proportion, dimensionless; is a frequency band index set with a frequency greater than Hz; is a frequency band index.

[0140] The calculation formula of the condition number is as follows:

[0141] ;

[0142] wherein, is the condition number, dimensionless; is the maximum singular value of the micro-hour frequency feature matrix, with the unit of millivolt; is the minimum singular value of the micro-hour frequency feature matrix, with the unit of millivolt.

[0143] The calculation formula of the variance of the transient stationary comprehensive feature vector is as follows:

[0144]

[0145] In the formula, is the variance of the transient stationary comprehensive feature vector, dimensionless; is the mean value of each element of the transient stationary comprehensive feature vector, dimensionless, and the calculation formula is .

[0146] The Lagrange function of the multi-parameter coupling optimization adjustment is as follows:

[0147]

[0148] In the formula, is the Lagrange function, dimensionless; is the sensor gain coefficient adjustment amount, with the unit of millivolt per Pascal; is the filter cutoff frequency adjustment amount, with the unit of kilohertz; is the Euclidean distance threshold adjustment amount, dimensionless; is the Lagrange multiplier, dimensionless; is the target value of the proportion of high-frequency components, taking a value of 0.6, dimensionless; is the target value of the condition number, taking a value of 15, dimensionless; is the target value of the variance, taking a value of 0.08, dimensionless; is the constraint condition function, dimensionless.

[0149] The specific implementation of step S09 is to construct a final electrical characteristic signal description vector, and the calculation formula of the intensity normalization value is as follows:

[0150]

[0151] In the formula, is the intensity normalization value, dimensionless; is the discharge intensity parameter, with the unit of decibel; is the maximum discharge intensity reference value, with the unit of decibel, which is determined according to the rated voltage level of the power equipment.

[0152] The construction formula of the final electrical characteristic signal description vector is as follows:

[0153]

[0154] In the formula,​​​​ is the dimension of the final electrical characteristic signal description vector, usually 20 to 40. is the type code value, 1 for internal discharge, 2 for surface discharge, 3 for corona discharge, dimensionless; is the effective feature component vector composed of elements with variance greater than 0.01 in the quasi-stationary comprehensive feature vector; is the statistical mean of the instantaneous frequency parameter, with the unit of hertz; is the statistical variance of the instantaneous frequency parameter, with the unit of hertz square; is the statistical mean of the instantaneous amplitude parameter, with the unit of millivolt; is the statistical variance of the instantaneous amplitude parameter, with the unit of millivolt square; is the statistical mean of the instantaneous phase parameter, with the unit of radian; is the statistical variance of the instantaneous phase parameter, with the unit of radian square; is the dimension of the final electrical characteristic signal description vector, usually 20 to 40.

[0155] The number of separable attention mechanism heads of the adaptive time-frequency analysis model is calculated according to the following formula:

[0156]

[0157] In the formula, is the number of separable attention mechanism heads, dimensionless; is the number of non-zero elements of the super-sparse frequency band energy distribution vector; is the rank of the micro-time-frequency feature matrix, dimensionless; is the effective dimension of the quasi-stationary comprehensive feature vector; is the normalized weight coefficient, dimensionless, satisfying , and the empirical values are 0.4, 0.3, and 0.3, respectively; the denominator 32 is the normalized reference value of the number of non-zero elements; is the normalized reference value of the total number of time sampling points; the denominator 40 is the normalized reference value of the effective dimension; represents the upward rounding operation; the reference head number is 8.

[0158] The calculation formula of the gating fusion unit of the separable attention mechanism is as follows:

[0159]

[0160]

[0161]

[0162] In the formula, is the output feature vector after fusion;​​​​​ This is the output feature vector of the spatial attention module; This is the output feature vector of the channel attention module; The sigmoid activation function is represented as follows: ; The gating weights for the spatial attention branch are dimensionless. The gating weights for the channel attention branches are dimensionless. This is the weight matrix of the gated unit; The bias vector of the gated unit; This means concatenating two vectors into one vector; is the base of the natural logarithm, with a value of approximately 2.71828; This is the input variable for the Sigmoid activation function.

[0163] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team is conducting partial discharge monitoring on a 220kV high-voltage transformer in a substation. This transformer has been in operation for 15 years and recently exhibited abnormal sounds, requiring precise extraction of electrical characteristic signals to determine the type and severity of the discharge. The technical team has deployed an 8-channel ultrasonic sensor array on the surface of the transformer tank. The sensors are broadband piezoelectric ceramic sensors with a response frequency range of 20 to 300 Hz. The sampling frequency is set to 2. Each data collection session lasts 100 seconds. After 72 hours of continuous monitoring, a total of 437 suspected discharge signals were collected, including a large amount of background noise from transformer cooling fans, oil pump vibrations, and the operation of surrounding high-voltage switches.

[0164] After acquiring the original ultrasonic signal sequence, the technical team first performed adaptive noise cancellation processing on the signal. The 8th channel, furthest from the transformer body, was selected as the reference noise channel, and the remaining 7 channels were used as the main signal channels. A noise estimation model was established using a minimum mean square error iterative algorithm, with an iteration step size of 0.005. The convergence criterion was that the error change after 50 consecutive iterations was less than [a certain value]. After an average of 183 iterations, the signal-to-noise ratio of each main signal channel increased from the initial 4.2. Increased to 18.7 The power of the background noise component was reduced by 85%.

[0165] The denoised ultrasonic signal sequence was subjected to a five-level wavelet packet decomposition using the Daubechies 8th-order wavelet basis function. The decomposition yielded 32 frequency band sub-signals with a bandwidth of 31.25. The band energy values of each frequency band sub-signal are calculated, as shown in Table 1.

[0166] Table 1 Energy distribution of partial frequency band sub-signals

[0167]

[0168] As can be seen from Table 1, the energy is mainly concentrated in the high frequency band of 156 to 281 , and the energy proportion of the first 5 frequency bands reaches 91.5%. According to the construction rule of the super-sparse frequency band energy distribution vector, the frequency bands with an energy proportion greater than 5% are selected as effective energy components, and the energy values of the remaining 26 frequency bands are set to zero, forming a super-sparse frequency band energy distribution vector with a sparsity of 81.3%. The proportion of high frequency components is calculated, and the sum of the frequency band energy greater than 100 is 0.958, accounting for 95.8%, far exceeding the threshold of 60%.

[0169] The technical team calls the adaptive time-frequency analysis model to perform time-frequency analysis on each frequency band sub-signal. The model adopts an encoder-decoder architecture, and the encoder includes four one-dimensional convolution layers with convolution kernel sizes of 64, 128, 256, and 512, respectively, and a step size of 2. The decoder adopts a separable attention mechanism, and according to the number of non-zero elements 6 of the super-sparse frequency band energy distribution vector, the rank 42 of the micro-time-frequency feature matrix, and the effective dimension 28 of the quasi-stationary comprehensive feature vector, the number of attention heads is calculated to be 7. As shown in Figure 2 , the empirical mode decomposition decomposes the denoising ultrasonic signal sequence into 9 intrinsic mode function components, and extracts the instantaneous frequency parameter, instantaneous amplitude parameter and instantaneous phase parameter after Hilbert transform of each component. The scale of the constructed micro-time-frequency feature matrix is 200000x512, of which the non-zero elements only account for 12.3%, and the sparsity reaches 87.7%.

[0170] Independent component analysis is performed on the micro-time-frequency feature matrix, and the FastICA algorithm is used to realize signal source separation. After 67 times of Newton iteration, 5 independent source signal components are separated. As shown in Figure 3 , the kurtosis value and skewness value of each independent source signal component are calculated, the kurtosis value of the first component is 8.73, and the skewness value is 2.14, which presents a significant sharp right-skewed distribution characteristic, which is consistent with the statistical characteristics of typical internal discharge signals. The kurtosis value of the second component is 2.87, and the skewness value is -0.36, which presents a flat left-skewed distribution, and is determined as residual background noise. The kurtosis values of the remaining 3 components are 4.52, 3.98 and 5.21, respectively.

[0171] The kurtosis value and skewness value of each independent source signal component are fused with the super-sparse frequency band energy distribution vector to construct a quasi-stationary comprehensive feature vector. The 20 The statistical characteristic parameters in the short time window include a root mean square value 0.342 , a peak factor 4.67, a waveform factor 1.89, a margin factor 12.34, a skewness value 2.14, and a kurtosis value 8.73. Six non-zero elements in the super-sparse frequency band energy distribution vector are extracted as frequency domain feature components. Energy barycentric positions of the micro-time-frequency feature matrix are calculated in the time axis direction and the frequency axis direction, and are 47.3 and 203.6 , respectively. Linear combination is performed according to weight coefficients 0.4, 0.35, and 0.25 to form a 53-dimensional quasi-stationary comprehensive feature vector. The variance of the vector is 0.063.

[0172] The technical team calculates the Euclidean distance between the quasi-stationary comprehensive feature vector and each type of standard discharge feature vector in the standard discharge feature vector library. The Euclidean distance of the internal discharge standard vector is 2.87, the Euclidean distance of the surface discharge standard vector is 8.43, and the Euclidean distance of the corona discharge standard vector is 11.26. The Euclidean distance threshold is set to 5.0, and the minimum distance 2.87 is less than the threshold, which is determined as an internal discharge signal. The signal duration is 1.3 , which is within the range of 0.1 to 2.0 , which meets the characteristics of the pulse type discharge mode, and a time domain peak value tracking prediction algorithm is used for analysis. Seventeen local maximum points in the signal are identified, the average time interval between adjacent peak points is 0.076 , and the amplitude attenuation ratio is 0.89.

[0173] At this time, the high-frequency component ratio 95.8% is greater than 60%, the condition number 12.8 of the micro-time-frequency feature matrix is less than 15, and the variance 0.063 of the quasi-stationary comprehensive feature vector is less than 0.08, all three conditions are met, and the multi-parameter coupling optimization adjustment mode is started. The Lagrange multiplier method is used to establish a constrained optimization model, and the optimal adjustment amount is obtained by 18 iterations. The sensor gain coefficient is reduced from the original 45 to 32 , to avoid signal saturation caused by high-frequency components. The lower cutoff frequency of the band-pass filter is adjusted from 80 to 120 , and the upper cutoff frequency is adjusted from 320 to 280 , to narrow the passband range to suppress high-frequency noise. The Euclidean distance threshold is adjusted from 5.0 to 4.2 to improve the sensitivity of discharge identification. After optimization and adjustment, the high-frequency component ratio is reduced to 68.4%, the condition number is increased to 13.7, and the variance is decreased to 0.052, and the three parameters reach a balanced state.

[0174] According to the identification result, the technical team constructs the final electrical characteristic signal description vector. The discharge type is internal discharge, and the type code value is 1. The discharge intensity parameter is 87.3 , the maximum discharge intensity reference value is set to 200 , and the intensity normalization value is calculated as 0.437. The effective feature components with a variance greater than 0.01 in the quasi-stable comprehensive feature vector are extracted, and a total of 31 effective components are selected. The statistical mean of the instantaneous frequency parameter is 198.7 , and the statistical variance is 23.4 . The statistical mean of the instantaneous amplitude parameter is 0.342 , and the statistical variance is 0.089 . The statistical mean of the instantaneous phase parameter is 1.67 , and the statistical variance is 0.43 . The type code value, intensity normalization value, 31 effective feature components, frequency statistical features, amplitude statistical features, and phase statistical features are arranged in order to form a 41-dimensional final electrical characteristic signal description vector, which is used for subsequent fault diagnosis and state evaluation.

[0175] The present application brings significant technical progress compared to traditional partial discharge detection methods. The traditional method usually uses fixed threshold energy detection or simple spectrum analysis, which is difficult to accurately extract discharge features in a strong background noise environment, and cannot effectively distinguish different types of discharge signals. The present application establishes a noise estimation model through adaptive noise cancellation processing, dynamically adjusts the filter parameters using the correlation between the reference channel and the main channel, and fundamentally improves the noise suppression capability. The construction of the super-sparse frequency band energy distribution vector realizes the precise positioning of the frequency domain features through wavelet packet decomposition and energy screening, avoiding the interference of redundant information caused by traditional full-band analysis. The adaptive time-frequency analysis model uses a separable attention mechanism to independently calculate and fuse the dependence relationship between the spatial dimension and the channel dimension, while maintaining the long-range modeling capability and significantly reducing the computational complexity, so that the model can run in real time on an embedded platform. Independent component analysis maximizes non-Gaussianity to separate signal source signals, effectively separating independent source signal components generated by different physical mechanisms, providing a reliable basis for accurate identification of discharge types. The quasi-stable comprehensive feature vector combines multi-dimensional features in the time domain, frequency domain, and time-frequency domain, fully describing the statistical properties and dynamic evolution law of the discharge signal. The multi-parameter coupled optimal adjustment mode adjusts the sensor gain, filter parameters, and identification threshold through a constraint optimization algorithm, realizes the adaptive optimization of system parameters, and significantly improves the detection stability and accuracy in complex working conditions.

[0176] It should be noted that the variables involved in the present application are explained in detail in Tables 2 and 3.

[0177] Table 2 Variable explanation table (first part)

[0178]

[0179] Table 3 Variable explanation table (second part)

[0180]

[0181] The above description is merely a specific implementation of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for extracting electrical characteristic signals of power equipment under complex operating conditions, characterized in that, A multi-channel ultrasonic sensor array is deployed on the surface of power equipment to collect ultrasonic signals generated by partial discharge, establishing an original ultrasonic signal sequence. Adaptive noise cancellation processing is applied to the original ultrasonic signal sequence to obtain a denoised ultrasonic signal sequence. Wavelet packet five-level decomposition is performed on the denoised ultrasonic signal sequence to obtain frequency band sub-signals of different frequency bands, and an ultra-sparse frequency band energy distribution vector is constructed. An adaptive time-frequency analytical model is used to perform time-frequency analysis on each frequency band sub-signal to extract instantaneous frequency parameters, instantaneous amplitude parameters, and instantaneous phase parameters, constructing a micro-hourly frequency characteristic matrix. Independent component analysis is performed on the micro-hourly frequency characteristic matrix to obtain independent sources. The signal components are analyzed and kurtosis and skewness values ​​are calculated. The kurtosis and skewness values ​​are then fused with the ultra-sparse frequency band energy distribution vector to construct a metastable comprehensive feature vector. The Euclidean distance between the metastable comprehensive feature vector and the standard discharge feature vectors of various types in the standard discharge feature vector library is calculated to determine the partial discharge signal and record the discharge type and discharge intensity parameters. The corresponding prediction algorithm is selected according to the discharge mode. Under certain conditions, the multi-parameter coupling optimization adjustment mode is activated. The final electrical characteristic signal description vector is constructed based on the discharge type, discharge intensity parameters, metastable comprehensive feature vector, instantaneous frequency parameters, instantaneous amplitude parameters, and instantaneous phase parameters.

2. The method according to claim 1, characterized in that, The adaptive noise cancellation process specifically involves establishing a noise estimation model using the reference noise channel signal and the main signal channel signal, and eliminating the background noise component through a minimum mean square error iterative algorithm.

3. The method according to claim 2, characterized in that, The minimum mean square error iterative algorithm specifically establishes a linear relationship model between the reference noise channel signal and the background noise component in the main signal channel signal, and uses the gradient descent method to iteratively update the filter coefficients to minimize the sum of squared errors between the filter output and the background noise component in the main signal channel signal.

4. The method according to claim 3, characterized in that, The five-level decomposition of the wavelet packet uses the Daubechies wavelet basis function, with a decomposition level of 5. After each level of decomposition, the high-frequency sub-signals and low-frequency sub-signals are further decomposed using the Mallat fast algorithm.

5. The method according to claim 4, characterized in that, The establishment of the ultra-sparse frequency band energy distribution vector specifically involves decomposing the wavelet packet into its fifth layer. Layer Each frequency band sub-signal is denoted as The frequency band energy value of the frequency band sub-signal is calculated as the sum of the squares of the amplitudes at each sampling point of the frequency band sub-signal. After the frequency band energy values ​​of all frequency band sub-signals in the layer are normalized, an energy distribution sequence is formed. The frequency band energy values ​​corresponding to the frequency bands with frequency band energy values ​​greater than 5% of the total energy are selected as the effective energy components, and the frequency band energy values ​​of the remaining frequency bands are set to zero.

6. The method according to claim 5, characterized in that, The adaptive time-frequency analytical model adopts an encoder-decoder architecture.

7. The method according to claim 6, characterized in that, The training of the adaptive time-frequency analytical model adopts a robust enhancement framework based on noise injection training. Training noise following a zero-mean Gaussian distribution is injected into the output feature map of each layer of the encoder, and the intensity of the training noise gradually increases from the first layer to the fourth layer.

8. The method according to claim 7, characterized in that, The establishment of the micro-hour frequency characteristic matrix is ​​specifically achieved by performing empirical mode decomposition on the noise-reduced ultrasonic signal sequence to obtain intrinsic mode function components, performing Hilbert transform on each intrinsic mode function component to obtain an analytical signal, calculating the instantaneous amplitude parameter and instantaneous frequency parameter of the analytical signal, and establishing a matrix on the time-frequency plane with time as the row index and frequency as the column index.

9. The method according to claim 8, characterized in that, The independent component analysis employs the FastICA algorithm, which achieves signal source separation by maximizing the non-Gaussianity of each independent source signal component. The non-Gaussianity metric is calculated using the negative entropy approximation method, and the Newton iteration method is used to accelerate convergence during the iteration process.

10. The method according to claim 9, characterized in that, The kurtosis is the ratio of the fourth central moment of the amplitude distribution of the independent source signal components to the square of the second central moment, and the skewness is the ratio of the third central moment of the amplitude distribution of the independent source signal components to the 1.5th power of the second central moment.

Images