Method for identifying and locating partial discharge of high-voltage cable

Abstract

The present invention proposes a method for identifying and locating partial discharges in high-voltage cables, which relates to the technical field of high-voltage cable diagnosis, and includes: collecting partial discharge pulse signals through high-frequency current transformers arranged at both ends of the high-voltage cable; using the adaptive multi-scale wavelet decomposition and reconstruction method to denoise the collected partial discharge pulse signals to obtain the denoised partial discharge characteristic signals; inputting the partial discharge characteristic signals into a pre-trained lightweight residual attention network for identification and classification, and outputting the discharge type discrimination result, discharge intensity, and characteristic parameters; based on the discharge type discrimination result, discharge intensity, and characteristic parameters, using a self-correcting double-end positioning algorithm to locate the discharge source and calculate the position of the discharge source. The present invention can solve the technical problems in the prior art such as unsatisfactory signal denoising effect, low accuracy of discharge type identification, and unstable positioning accuracy, so as to achieve high-precision identification and location of partial discharges in high-voltage cables.

Description

Technical Field

[0001] The present invention relates to the technical field of high-voltage cable diagnosis, and particularly to a method for identifying and locating partial discharges in high-voltage cables. Background Art

[0002] High-voltage cables are key equipment in the power system, and their safe operation is directly related to the reliability of power supply. Partial discharge is an important manifestation of the deterioration of the insulation of high-voltage cables, and can reflect the type, location and development degree of insulation defects. Therefore, accurately identifying the type of partial discharge and achieving precise positioning are of great significance for preventing cable faults and ensuring the safe operation of the power grid.

[0003] Existing partial discharge identification technologies are mainly based on signal processing and pattern recognition methods. In terms of signal processing, methods such as wavelet transform and empirical mode decomposition are used for signal denoising and feature extraction; in terms of discharge type identification, pattern recognition is mainly carried out based on phase spectrum features, pulse waveform features, etc., such as clustering analysis based on statistical parameters, feature extraction based on phase distribution diagrams, etc. In terms of the location of the discharge source, single-end location methods and double-end location methods are mainly used. Among them, the double-end location method calculates the location of the discharge point by measuring the time difference between the arrival of the discharge signal at both ends of the cable, and has a high location accuracy.

[0004] However, the existing technologies still have the following problems: First, the denoising effect of traditional signal processing methods is not ideal in complex electromagnetic environments, and the feature extraction ability is limited; second, the identification methods based on fixed features are difficult to adapt to the differences of different types of discharge signals, and the identification accuracy is limited; third, existing location algorithms often ignore the influence of cable body parameters, environmental factors, etc. on the signal propagation characteristics in practical applications, resulting in unstable location accuracy. In addition, due to the lack of systematic signal processing and analysis methods, it is difficult to achieve accurate extraction and reliable judgment of discharge characteristics, which restricts the further development of detection technologies. Summary of the Invention

[0005] In view of this, the present invention proposes a method for identifying and locating partial discharges in high-voltage cables. By using the method of adaptive multi-scale wavelet decomposition and reconstruction to denoise the discharge signal, combining a lightweight residual attention network to achieve accurate identification of the discharge type, and based on a self-calibrating double-end location algorithm to achieve accurate location of the discharge source, so as to solve the technical problems of unsatisfactory signal denoising effect, low discharge type identification accuracy, unstable location accuracy, etc. in the existing technologies, thereby realizing high-precision identification and location of partial discharges in high-voltage cables.

[0006] The technical solution of the present invention is realized as follows:

[0007] The present invention provides a method for identifying and locating partial discharges in high-voltage cables, including:

[0008] S1. Collect partial discharge pulse signals through high-frequency current transformers arranged at both ends of the high-voltage cable;

[0009] S2. Use the adaptive multi-scale wavelet decomposition and reconstruction method to denoise the collected partial discharge pulse signals to obtain the denoised partial discharge characteristic signals;

[0010] S3. Input the partial discharge characteristic signals into a pre-trained lightweight residual attention network for identification and classification. The lightweight residual attention network includes a parallel multi-scale decomposition module, a frequency-domain guided attention module, a feature adaptive fusion module, and an output module, and outputs the discharge type discrimination result, discharge intensity, and characteristic parameters;

[0011] S4. Based on the discharge type discrimination result, discharge intensity, and characteristic parameters, use the self-corrected double-end positioning algorithm to locate the discharge source. By measuring the time difference of the partial discharge pulse signals arriving at both ends of the cable and combining the cable length parameters, calculate the position of the discharge source.

[0012] On the basis of the above technical solution, preferably, step S2 includes:

[0013] S21. Select wavelet basis functions for the collected partial discharge pulse signals;

[0014] S22. Perform n-layer wavelet decomposition on the partial discharge pulse signals to obtain the approximation coefficient a n and the detail coefficients d1 - d n ;

[0015] S23. Calculate the signal-to-noise ratio SNR j :

[0016]

[0017] In the formula, is the signal variance, is the noise variance, j = 1, 2,..., n represents the decomposition layer;

[0018] S24. Establish an adaptive weight coefficient calculation model:

[0019]

[0020] In the formula, w j is the adaptive weight coefficient;

[0021] S25. Based on the adaptive weight coefficient and the corresponding layer of detail coefficients, perform wavelet reconstruction according to the reconstruction signal model to obtain the denoised partial discharge characteristic signals, where the reconstruction signal model is:

[0022]

[0023] Wherein, Y(t) represents the reconstructed signal, and T j (d j ) represents the threshold function.

[0024] Based on the above technical solution, preferably, the threshold function T j (d j ) adopts the soft threshold function:

[0025] T j (d j ) = sign(d j )(|d j | - λ j ), |d j | > λ j

[0026] T j (d j ) = 0, |d j | ≤ λ j

[0027] Wherein, λ j is the adaptive threshold of the j-th layer:

[0028] λ j = σ j ·(2lnN) 0.5

[0029] Wherein, σ j is the noise standard deviation of the j-th layer, and N is the signal length.

[0030] Based on the above technical solution, preferably, step S2 further includes: determining the optimal decomposition scale n based on the signal frequency characteristics of the partial discharge pulse signal:

[0031]

[0032] Wherein, f s is the sampling frequency, f min is the lowest frequency of interest, and floor is the floor function.

[0033] Based on the above technical solution, preferably, the structure of the lightweight residual attention network includes:

[0034] The parallel multi-scale decomposition module contains three parallel branches, each branch consisting of two depthwise separable convolutional residual blocks. Among them, the first branch is set with a 3×3 convolutional kernel, a stride of 1, and 64 channels; the second branch is set with a 5×5 convolutional kernel, a stride of 1, and 32 channels; the third branch is set with a 7×7 convolutional kernel, a stride of 1, and 16 channels. Each residual block contains two layers of depthwise separable convolution and a skip connection.

[0035] The frequency-domain guided attention module contains an input layer, two fully connected layers, and an attention weight calculation layer.

[0036] The feature adaptive fusion module includes: an energy distribution calculation unit that calculates the energy distribution of the feature maps of each branch; an adaptive weight generation unit that outputs the fusion weights of each branch; and a feature fusion unit that performs a concatenate operation on the weighted features.

[0037] The output module includes three task branches, namely the discharge type branch, the discharge intensity branch, and the feature parameter branch. Each task branch contains two layers of fully connected layers.

[0038] On the basis of the above technical solutions, preferably, the process of the lightweight residual attention network for processing partial discharge feature signals is as follows:

[0039] Input the partial discharge feature signal into the parallel multi-scale decomposition module. The three parallel branches simultaneously process the input partial discharge feature signal, and each branch outputs a feature map of the corresponding scale.

[0040] Input the feature maps output by the parallel multi-scale decomposition module into the frequency-domain guided attention module. Calculate the spectral features through the fast Fourier transform in the input layer, extract the frequency-domain features through two fully connected layers, and the output dimensions are 128 and 64 respectively. Use the attention weight calculation layer to generate the attention weights and output a weight matrix with the same dimension as the input feature map. Weight the feature map according to the weight matrix.

[0041] Input the weighted feature maps into the feature adaptive fusion module, calculate the feature energy of the feature maps of each branch, calculate the fusion weights based on the energy distribution, and perform weighted fusion to obtain the final feature representation.

[0042] Use the final feature representation as the shared feature and input it into the output module, which is processed through three task branches respectively:

[0043] Discharge type branch: Input the shared feature into the first fully connected layer with 128 neurons, use ReLU activation, and then input it into the second fully connected layer with the same number as the number of discharge types, use softmax activation, output the probability distribution of each discharge type, and select the category with the highest probability as the discharge type discrimination result.

[0044] Discharge intensity branch: The shared features are input into the first fully connected layer with 128 neurons, activated by ReLU, and then input into the second fully connected layer with a single neuron, and the intensity value is output using a linear activation function as the discharge intensity;

[0045] Feature parameter branch: The shared features are input into the first fully connected layer with 128 neurons, activated by ReLU, and then input into the second fully connected layer with the same number of neurons as the feature parameters, and the parameter values are output using a linear activation function to obtain the feature parameters.

[0046] Based on the above technical solution, preferably, step S4 includes:

[0047] S41. According to the discharge type discrimination result, select the corresponding propagation speed correction coefficient, and combine the discharge intensity coefficient to correct the propagation speed;

[0048] S42. Use the main frequency feature, amplitude feature and waveform feature of the signal to establish a signal propagation model considering multiple factors, and calculate the comprehensive correction term;

[0049] S43. Through the GPS high-precision time synchronization system, collect the timestamps when the signal arrives at both ends of the cable, calculate the maximum value point of the cross-correlation function of the signals at both ends, and obtain the accurate time difference;

[0050] S44. Combine the cable length parameter, the corrected propagation speed and the accurate time difference to calculate the position of the discharge source.

[0051] Based on the above technical solution, preferably, the correction formula for the propagation speed is:

[0052] v = v0·(1 + k1·type + k2·strength)

[0053] In the formula, v is the corrected propagation speed, v0 is the theoretical propagation speed in the high-voltage cable, k1 is the discharge type correction coefficient, k2 is the discharge intensity correction coefficient, type is the discharge type coefficient, and strength is the discharge intensity coefficient;

[0054] The formula for the accurate time difference is:

[0055] τ = argmax{R 12 (t)}

[0056] In the formula, τ is the accurate time difference, R 12 (t) is the cross-correlation function of the signals at both ends, R 12R(t) = ∫x1(τ)x2(τ - t)dτ, where x1(τ) is the signal collected at the first end, x2(τ) is the signal collected at the second end, t is the time delay value to be attempted, and x2(τ - t) represents shifting the second-end signal to the right by t time units;

[0057] The formula for the comprehensive correction term is:

[0058] δ = α·freq + β·amp + γ·shape

[0059] In the formula, δ is the comprehensive correction term, freq is the main frequency feature of the signal, amp is the amplitude feature of the signal, shape is the waveform feature of the signal, and α, β, γ are the weight coefficients of the corresponding features;

[0060] The formula for the position of the discharge source is:

[0061]

[0062] In the formula, d is the distance from the partial discharge point to the reference end, and L is the total length of the cable.

[0063] Based on the above technical solution, preferably, the discrimination results of the discharge types include three categories, namely internal discharge, surface discharge, and corona discharge. Correspondingly, the values of the discharge type coefficient type are:

[0064] Internal discharge: type = 1.0; Surface discharge: type = 0.8; Corona discharge: type = 0.6;

[0065] The formula for the discharge intensity coefficient strength is:

[0066] strength = min(1.0, q / q0);

[0067] In the formula, q is the actual discharge intensity, and q0 is the theoretical discharge intensity threshold.

[0068] Based on the above technical solution, preferably, the sampling accuracy of the high-frequency current transformer is 14 bits, the sampling frequency is 10 - 100 MS / s, and the bandwidth is 0.01 - 10 MHz.

[0069] The present invention has the following beneficial effects compared with the prior art:

[0070] (1) The present invention constructs a complete technical system for the identification and location of partial discharges in high-voltage cables. High-precision signal acquisition is achieved through high-frequency current transformers, and the organic integration of adaptive multi-scale wavelet decomposition for noise reduction, lightweight residual attention network identification, and self-calibration dual-end location algorithm is realized. This technical system realizes the full-process automatic processing from signal acquisition, feature extraction to type identification and location, improving the recognition accuracy and location precision of partial discharge detection;

[0071] (2) The adaptive multi-scale wavelet decomposition and reconstruction method of the present invention realizes the dynamic optimization of decomposition coefficients in different frequency bands by introducing an adaptive weight calculation model based on signal-to-noise ratio. This method can automatically adjust the weight coefficients of each scale according to signal characteristics, effectively suppressing background noise while maintaining the integrity of the effective signal;

[0072] (3) The lightweight residual attention network designed by the present invention adopts a combined structure of a parallel multi-scale decomposition module, a frequency-domain guided attention module, and a feature adaptive fusion module. Through multi-scale feature extraction, frequency-domain energy distribution analysis, and feature dynamic fusion, it realizes the accurate capture and classification of the characteristics of different types of discharge signals. While reducing the computational complexity, this network structure effectively improves the accuracy of discharge type recognition;

[0073] (4) The self-calibration dual-end location algorithm of the present invention establishes a signal propagation model with multi-factor correction. By introducing a discharge type coefficient and a discharge intensity coefficient, and combining the comprehensive influence of signal characteristic parameters, it realizes the adaptive correction of the propagation speed. Cooperating with the GPS high-precision time synchronization system, this algorithm effectively improves the accuracy of the discharge source location, providing a reliable location basis for fault repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0075] Figure 1 is the flowchart of the method of the present invention;

[0076] Figure 2 is the network structure diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0078] As Figure 1 shown, the present invention provides a method for identifying and locating partial discharges in high-voltage cables, including:

[0079] S1. Collect partial discharge pulse signals through high-frequency current transformers arranged at both ends of the high-voltage cable;

[0080] S2. Use the adaptive multi-scale wavelet decomposition and reconstruction method to perform noise reduction processing on the collected partial discharge pulse signals to obtain the denoised partial discharge characteristic signals;

[0081] S3. Input the partial discharge characteristic signals into a pre-trained lightweight residual attention network for identification and classification. The lightweight residual attention network includes a parallel multi-scale decomposition module, a frequency-domain guided attention module, a feature adaptive fusion module, and an output module, and outputs the discharge type discrimination result, discharge intensity, and characteristic parameters;

[0082] S4. Based on the discharge type discrimination result, discharge intensity, and characteristic parameters, use the self-corrected double-end positioning algorithm to locate the discharge source. By measuring the time difference between the arrival of the partial discharge pulse signal at both ends of the cable and combining the cable length parameter, calculate the position of the discharge source.

[0083] The method for identifying and locating partial discharge of high-voltage cables provided by the present invention first collects partial discharge pulse signals through high-frequency current transformers arranged at both ends of the high-voltage cable; subsequently, the adaptive multi-scale wavelet decomposition and reconstruction method is used to denoise the collected signals. This method selects an appropriate wavelet basis function, performs n-layer wavelet decomposition on the signal to obtain approximation coefficients and detail coefficients, establishes an adaptive weight coefficient model based on signal-to-noise ratio calculation, and uses a soft threshold function for signal reconstruction to obtain the denoised partial discharge characteristic signal; then, the denoised characteristic signal is input into a pre-trained lightweight residual attention network for identification and classification. This network includes a parallel multi-scale decomposition module (composed of three parallel branches with different convolution kernel sizes), a frequency-domain guided attention module (including an input layer, a fully connected layer, and an attention weight calculation layer), and a feature adaptive fusion module (including an energy distribution calculation unit, an adaptive weight generation unit, and a feature fusion unit), and outputs the discharge type discrimination result (internal discharge, surface discharge, and corona discharge), discharge intensity, and characteristic parameters through three task branches; finally, based on the recognition result, a self-correcting double-end positioning algorithm is used to locate the discharge source. This algorithm considers the comprehensive influence of the discharge type coefficient, discharge intensity coefficient, and signal characteristics (main frequency, amplitude, waveform), combines the GPS high-precision time synchronization system to collect the timestamps of the signals arriving at both ends, calculates the maximum point of the cross-correlation function of the two-end signals to obtain the accurate time difference, and combines the cable length parameter and the corrected propagation speed to calculate the discharge source position, realizing the accurate positioning of partial discharge.

[0084] Specifically, in an embodiment of the present invention, step S1 includes:

[0085] A high-frequency current transformer is installed at each end of the high-voltage cable to collect high-frequency pulse current signals generated by partial discharge. The sampling accuracy of the high-frequency current transformer is 14 bits, the sampling frequency can be adjusted in the range of 10 - 100 MS / s, and the bandwidth is 0.01 - 10 MHz, which can meet the high-frequency sampling requirements of partial discharge pulse signals.

[0086] Specifically, the high-frequency current transformer adopts a split structure, which is convenient for installation at the grounding coils at both ends of the cable. During installation, the primary coil of the high-frequency current transformer is sleeved on the grounding wire of the cable, and the secondary coil is connected to the data acquisition unit through a coaxial cable. The acquisition unit digitally converts the collected analog signal through a high-speed AD converter and transmits the digital signal to the data processing unit.

[0087] In practical applications, a high-frequency current transformer simultaneously collects partial discharge signals at both ends of a cable. The collected signals contain characteristic information such as the amplitude, phase, and waveform of the discharge pulses. Since partial discharge signals have the characteristics of high frequency and narrow pulses, a high sampling rate and high-precision sampling method can accurately capture the fast-changing characteristics of the discharge pulses, providing reliable raw data for subsequent signal processing and analysis.

[0088] Specifically, in an embodiment of the present invention, step S2 includes:

[0089] S21. Select a wavelet basis function for the collected partial discharge pulse signal; according to the characteristics of the partial discharge pulse signal, select a wavelet basis function with good time-frequency localization characteristics, such as orthogonal wavelet bases like db4 and sym4. These wavelet basis functions have compact support and symmetry, and are suitable for processing the transient characteristics in partial discharge signals.

[0090] S22. Perform n-layer wavelet decomposition on the partial discharge pulse signal to obtain the approximation coefficient a n and the detail coefficients d1 - d n ;

[0091] In this step, it also includes: determining the optimal decomposition scale n based on the signal frequency characteristics of the partial discharge pulse signal:

[0092]

[0093] where f s is the sampling frequency, f min is the lowest frequency of interest, and floor is the floor function. This method can ensure that the decomposition scale matches the frequency characteristics of the signal, improving the noise reduction effect.

[0094] Specifically, when performing wavelet decomposition, the decomposition process uses the pyramid algorithm to recursively decompose the signal through high-pass and low-pass filters, enabling the signal to be fully represented in different frequency bands.

[0095] S23. Calculate the signal-to-noise ratio SNR j of each layer of detail coefficients:

[0096]

[0097] where is the signal variance, is the noise variance, and j = 1, 2,..., n represents the decomposition layer number;

[0098] S24. Establish an adaptive weight coefficient calculation model:

[0099]

[0100] where w j is the adaptive weight coefficient; the model automatically assigns weights according to the signal-to-noise ratio of each layer of signals, so that the layer with a higher signal-to-noise ratio has a larger weight, thereby retaining more useful signal components.

[0101] S25. Based on the adaptive weight coefficient and the detail coefficients of the corresponding layer, perform wavelet reconstruction according to the reconstructed signal model to obtain the denoised partial discharge characteristic signal, where the reconstructed signal model is:

[0102] Y(t) = ∑(w j ·T j (d j ))

[0103] where Y(t) represents the reconstructed signal, and T j (d j ) represents the threshold function. The threshold function T j (d j ) adopts the soft threshold function:

[0104] T j (d j ) = sign(d j )(|d j | - λ j ), |d j | > λ j

[0105] T j (d j ) = 0, |d j | ≤ λ j

[0106] where λ j is the adaptive threshold of the j-th layer:

[0107] λ j = σ j ·(2lnN) 0.5

[0108] where σ j is the noise standard deviation of the j-th layer, and N is the signal length. The soft threshold function can smooth the signal and avoid the discontinuity brought by the hard threshold function.

[0109] The reconstruction process of this embodiment comprehensively considers the weights of each layer of signals and the threshold processing results, can effectively retain the discharge characteristics, and at the same time suppress the background noise.

[0110] Specifically, as Figure 2 shown, in an embodiment of the present invention, the structure of the lightweight residual attention network includes:

[0111] The parallel multi-scale decomposition module contains three parallel branches, each branch consisting of two depthwise separable convolutional residual blocks. Among them, the first branch is set with a 3×3 convolutional kernel, a stride of 1, and 64 channels, which is used to capture local fine features; the second branch is set with a 5×5 convolutional kernel, a stride of 1, and 32 channels, which is used to capture medium-scale features; the third branch is set with a 7×7 convolutional kernel, a stride of 1, and 16 channels, which is used to capture large-scale context features; each residual block contains two layers of depthwise separable convolution and a skip connection; it is used to capture feature information of different scales. Specifically, the internal structure of each residual block is set as follows:

[0112] The first layer of depthwise separable convolution: pointwise convolution + depthwise convolution, BN layer, ReLU activation;

[0113] The second layer of depthwise separable convolution: pointwise convolution + depthwise convolution, BN layer;

[0114] Residual connection: adding the input feature to the output of the second layer;

[0115] Output activation: ReLU function.

[0116] The frequency-domain guided attention module contains: an input layer that receives the feature maps output by the parallel branches; two fully connected layers and an attention weight calculation layer; it is used to extract and strengthen frequency-domain features. It can also include a frequency-domain transformation layer, which is set after the input layer and is used to perform a two-dimensional Fourier transform on the feature maps. Specifically, the attention weight calculation layer includes: a channel attention branch that generates attention weights in the channel dimension; a spatial attention branch that generates attention weights in the spatial dimension; attention fusion that combines channel attention and spatial attention.

[0117] The feature adaptive fusion module includes: an energy distribution calculation unit that calculates the energy distribution of the feature maps of each branch; an adaptive weight generation unit that outputs the fusion weights of each branch; a feature fusion unit that performs a concatenate operation on the weighted features;

[0118] The output module includes three task branches, namely the discharge type branch, the discharge intensity branch, and the feature parameter branch, and each task branch contains two layers of fully connected layers.

[0119] The process of the lightweight residual attention network for processing partial discharge feature signals is as follows:

[0120] Input the partial discharge feature signal into the parallel multi-scale decomposition module. The three parallel branches simultaneously process the input partial discharge feature signal, and each branch outputs the feature map corresponding to its scale; for the input signal x, the processing of each branch i can be expressed as: Among them, Denote the j-th depthwise separable convolutional residual block of the i-th branch, and three groups of feature maps with different scales {F1(x), F2(x), F3(x)} are obtained.

[0121] Input the feature maps output by the parallel multi-scale decomposition module into the frequency-domain guided attention module. Calculate the spectral feature S(F) = FFT(F) through the fast Fourier transform at the input layer. Extract the frequency-domain feature V = W2·ReLU(W1·S(F)) through two fully connected layers, with the output dimensions being 128 and 64 respectively. Use the attention weight calculation layer to generate the attention weight, and output the weight matrix A = sigmoid(V) with the same dimension as the input feature map. Weight the feature map according to the weight matrix: F' i (x) = F i (x)⊙A, where ⊙ represents element-wise multiplication;

[0122] Input the weighted feature map into the feature adaptive fusion module, and calculate the feature energy of each branch feature map: E i = ||F' i (x)||2. Calculate the fusion weight based on the energy distribution: w i = softmax(E i ). Perform weighted fusion to obtain the final feature representation: F * = concat(w1·F'1(x), w2·F'2(x), w3·F'3(x)).

[0123] Use the final feature representation as the shared feature and input it into the output module, which is processed through three task branches respectively:

[0124] Discharge type branch: Input the shared feature into the first fully connected layer with 128 neurons, use ReLU activation, and then input it into the second fully connected layer with the same number of neurons as the number of discharge types, use softmax activation, output the probability distribution of each discharge type, and select the category with the highest probability as the discharge type discrimination result;

[0125] Discharge intensity branch: Input the shared feature into the first fully connected layer with 128 neurons, use ReLU activation, and then input it into the second fully connected layer with a single neuron, use the linear activation function to output the intensity value as the discharge intensity;

[0126] Feature parameter branch: Input the shared feature into the first fully connected layer with 128 neurons, use ReLU activation, and then input it into the second fully connected layer with the same number of neurons as the number of feature parameters, use the linear activation function to output each parameter value to obtain the feature parameters.

[0127] Specifically, in this embodiment, the pre-training process of the lightweight residual attention network is as follows:

[0128] The network adopts an end-to-end training method, uses the Adam optimizer for parameter optimization, sets the initial learning rate to 0.001, and dynamically adjusts it using the cosine annealing strategy. To fully consider the particularity of partial discharge signals, an adaptive multi-task loss function based on frequency-domain energy distribution is designed:

[0129] L = L type + λ1·L strength + λ2·L param + λ3·L freq

[0130] where L type is the discharge type classification loss, and the cross-entropy loss with frequency-domain energy weights is adopted:

[0131] L type = -∑(y i · log(p i )· w i )

[0132]

[0133] In the formula, y i is the one-hot encoding of the true label, p i is the probability distribution predicted by the model, w i is the frequency-domain energy weight, and E i is the energy of the i-th type of sample in a specific frequency band.

[0134] L strength is the discharge intensity regression loss, considering the signal amplitude distribution characteristics:

[0135]

[0136] In the formula, y1 is the true discharge intensity value, is the predicted discharge intensity value, α is the frequency-domain difference weight coefficient, FFT(y), represents the Fourier transform.

[0137] L param is the feature parameter regression loss, introducing the correlation constraint between parameters:

[0138]

[0139] In the formula, y2 is the true feature parameter vector, is the predicted feature parameter vector, β is the correlation constraint weight, is the Pearson correlation coefficient between parameters.

[0140] L freqis the frequency-domain consistency loss, ensuring the performance of the model in the frequency domain:

[0141]

[0142] where PSD(y3) is the power spectral density of the true signal, is the power spectral density of the predicted signal, and KL is the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions.

[0143] λ1, λ2, and λ3 are the weight coefficients of each term in the loss function, with values in the range [0, 1], and the initial values are set to λ1 = 0.4, λ2 = 0.3, and λ3 = 0.3.

[0144] During training, a parameter dynamic adjustment strategy is adopted, as follows:

[0145] The initial value of the learning rate is 0.001, the cosine annealing period is 10 epochs, and the minimum learning rate is 10 -6 , and it restarts when the validation loss does not decrease for 3 consecutive epochs. Based on the performance of the validation set, the weight coefficients are updated after each epoch. If the decrease rate of a certain loss is significantly slower than other terms, the corresponding λ value increases; if a certain loss decreases too fast, which may lead to overfitting, the corresponding λ value decreases. The adjustment step size is ±0.05 to ensure that λ1 + λ2 + λ3 = 1.

[0146] After the lightweight residual attention network is trained, it can be deployed in an embedded processor, such as an ARM Cortex-A72 or a higher-performance processor. After collecting the partial discharge pulse signal using a high-frequency current transformer, noise reduction is performed, and the partial discharge characteristic signal is input into the trained lightweight residual attention network for identification and classification, and the discharge type discrimination result, discharge intensity, and characteristic parameters are output.

[0147] In this embodiment, the discharge type classification includes three types: internal discharge, surface discharge, and corona discharge. The discharge type branch outputs a probability distribution vector, which is activated using softmax to represent the possibility of each type. For example, internal discharge: 0.85 (85%); surface discharge: 0.10 (10%); corona discharge: 0.05 (5%); at this time, the network determines that the discharge signal is most likely an internal discharge, with a confidence of 85%.

[0148] In this embodiment, the discharge intensity branch outputs a single value, which is activated linearly to represent the discharge amount, with the unit of pC (picocoulomb). For example, the detected discharge intensity is 500 pC.

[0149] In this embodiment, the characteristic parameter branch outputs a multi-dimensional vector, which is linearly activated to represent various characteristic parameters of the discharge signal. The characteristic parameters include rise time, duration, waveform, main frequency, amplitude, etc. For example, rise time: 0.5 μs; duration: 2.5 μs; waveform characteristic: 0.75; main frequency: 2.5 MHz; amplitude characteristic: 0.85. These parameters are used to describe the time-domain and frequency-domain characteristics of the discharge pulse.

[0150] Specifically, in an embodiment of the present invention, step S4 includes:

[0151] S41. According to the discharge type discrimination result, select the corresponding propagation speed correction coefficient, and combine the discharge intensity coefficient to correct the propagation speed. The correction formula for the propagation speed is:

[0152] v = v0·(1 + k1·type + k2·strength)

[0153] In the formula, v is the corrected propagation speed, v0 is the theoretical propagation speed in the high-voltage cable, k1 is the discharge type correction coefficient, k2 is the discharge intensity correction coefficient, type is the discharge type coefficient, and strength is the discharge intensity coefficient. The value of the discharge type coefficient type is:

[0154] Internal discharge: type = 1.0; surface discharge: type = 0.8; corona discharge: type = 0.6;

[0155] The calculation formula for the discharge intensity coefficient strength is:

[0156] strength = min(1.0, q / q0);

[0157] In the formula, q is the actual discharge intensity, and q0 is the theoretical discharge intensity threshold.

[0158] Specifically, the value range of k1 is 0.1 - 0.3. Considering the influence degree of different discharge types on the propagation speed, the internal discharge has the greatest influence, and the corona discharge has the smallest influence. The value range of k2 is 0.2 - 0.4. Considering the influence degree of the discharge intensity on the propagation speed, it maintains a reasonable proportional relationship with k1.

[0159] S42. Use the main frequency characteristic, amplitude characteristic, and waveform characteristic of the signal to establish a signal propagation model considering the influence of multiple factors, and calculate the comprehensive correction term;

[0160] The calculation formula for the comprehensive correction term is:

[0161] δ = α·freq + β·amp + γ·shape

[0162] Where δ is the comprehensive correction term, freq is the main frequency feature of the signal, amp is the amplitude feature of the signal, shape is the waveform feature of the signal, and α, β, and γ are the weight coefficients of the corresponding features.

[0163] Specifically, since the main frequency feature has the greatest impact on the positioning accuracy and is relatively stable and reliable, while the amplitude feature is of secondary importance and is affected by signal attenuation, and the waveform feature is relatively unstable and is greatly affected by the propagation process. Therefore, when setting α, β, and γ, first use the recommended values for initial setting, and then fine-tune according to the actual test results, while maintaining the constraint of α + β + γ = 1, and α > β > γ.

[0164] S43. Through the GPS high-precision time synchronization system, collect the timestamps of the signal arriving at both ends of the cable, calculate the maximum point of the cross-correlation function of the two-end signals, and obtain the accurate time difference; the formula for calculating the accurate time difference is:

[0165] τ = argmax{R 12 (t)}

[0166] Where τ is the accurate time difference, and R 12 (t) is the cross-correlation function of the two-end signals, and R 12 (t) = ∫x1(τ)x2(τ - t)dτ, x1(τ) is the signal collected at the first end, x2(τ) is the signal collected at the second end, t is the time delay value to be tried, and x2(τ - t) represents moving the second-end signal to the right by t time units.

[0167] S44. Combine the cable length parameter, the corrected propagation speed, and the accurate time difference to calculate the position of the discharge source. The formula for calculating the position of the discharge source is:

[0168]

[0169] Where d is the distance from the partial discharge point to the reference end, and L is the total length of the cable.

[0170] The present invention adopts a dual-end high-precision positioning technology based on GPS time service. By installing high-frequency current transformers and high-speed sampling devices at both ends of the cable, and using the 3ns precision time synchronization signal provided by Beidou GPS, the partial discharge pulse signals are synchronously collected. The system first corrects the propagation speed according to the discharge type and discharge intensity; then calculates the comprehensive correction term based on the main frequency feature, amplitude feature, and waveform feature of the signal; then calculates the accurate time difference between the two-end signals through cross-correlation analysis; finally, combines the cable length parameter L to calculate the position of the discharge source, and the positioning accuracy can reach 0.2%L ± 5m. The whole process uses FPGA to cooperate with high-speed ADC for real-time data processing, and uploads the positioning results to the background server in real time through the 4G wireless network.

[0171] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for identifying and locating partial discharge in high-voltage cables, characterized in that, Including: S1. Collect partial discharge pulse signals through high-frequency current transformers arranged at both ends of high-voltage cables; S2. Use the adaptive multi-scale wavelet decomposition and reconstruction method to denoise the collected partial discharge pulse signals to obtain the denoised partial discharge characteristic signals; S3. Input the partial discharge characteristic signals into a pre-trained lightweight residual attention network for identification and classification. The lightweight residual attention network includes a parallel multi-scale decomposition module, a frequency-domain guided attention module, a feature adaptive fusion module, and an output module, and outputs the discharge type discrimination result, discharge intensity, and characteristic parameters; The structure of the lightweight residual attention network includes: The parallel multi-scale decomposition module contains three parallel branches, and each branch consists of two depthwise separable convolutional residual blocks. Among them, the first branch is set with a 3×3 convolutional kernel, a stride of 1, and 64 channels; the second branch is set with a 5×5 convolutional kernel, a stride of 1, and 32 channels; the third branch is set with a 7×7 convolutional kernel, a stride of 1, and 16 channels; each residual block contains two layers of depthwise separable convolution and a skip connection; The frequency-domain guided attention module contains an input layer, two fully connected layers, and an attention weight calculation layer; The feature adaptive fusion module includes: an energy distribution calculation unit that calculates the energy distribution of the feature maps of each branch; an adaptive weight generation unit that outputs the fusion weights of each branch; a feature fusion unit that performs a concatenate operation on the weighted features; The output module includes three task branches, namely the discharge type branch, the discharge intensity branch, and the characteristic parameter branch, and each task branch contains two layers of fully connected layers; The process of the lightweight residual attention network processing the partial discharge characteristic signals is as follows: Input the partial discharge characteristic signals into the parallel multi-scale decomposition module, and the three parallel branches simultaneously process the input partial discharge characteristic signals, and each branch outputs the feature maps corresponding to the scales; Input the feature maps output by the parallel multi-scale decomposition module into the frequency-domain guided attention module. Calculate the spectral features through the fast Fourier transform in the input layer, extract the frequency-domain features through two fully connected layers, and the output dimensions are 128 and 64 respectively. Use the attention weight calculation layer to generate the attention weights and output a weight matrix with the same dimension as the input feature maps, and weight the feature maps according to the weight matrix; Input the weighted feature maps into the feature adaptive fusion module, calculate the feature energy of the feature maps of each branch, calculate the fusion weights based on the energy distribution, and perform weighted fusion to obtain the final feature representation; Use the final feature representation as the shared feature and input it into the output module, which is processed through three task branches respectively: Discharge type branch: Input the shared feature into the first fully connected layer with 128 neurons, use ReLU activation, and then input it into the second fully connected layer with the same number as the discharge type quantity, use softmax activation, output the probability distribution of each discharge type, and select the category with the highest probability as the discharge type discrimination result; Discharge intensity branch: The shared features are input into the first fully-connected layer with 128 neurons, activated by ReLU, and then input into the second fully-connected layer with a single neuron, and the intensity value is output using a linear activation function as the discharge intensity. Feature parameter branch: The shared features are input into the first fully-connected layer with 128 neurons, activated by ReLU, and then input into the second fully-connected layer with the same number of neurons as the feature parameters, and the parameter values are output using a linear activation function to obtain the feature parameters. S4. Based on the discharge type discrimination result, discharge intensity, and feature parameters, use the self-correcting double-end positioning algorithm to locate the discharge source. By measuring the time difference of the partial discharge pulse signal arriving at both ends of the cable, and combining the cable length parameter, calculate the position of the discharge source. Step S4 includes: S41. According to the discharge type discrimination result, select the corresponding propagation speed correction coefficient, and combine the discharge intensity coefficient to correct the propagation speed. S42. Use the main frequency feature, amplitude feature, and waveform feature of the signal to establish a signal propagation model considering multiple factors, and calculate the comprehensive correction term. S43. Through the GPS high-precision time synchronization system, collect the timestamps of the signal arriving at both ends of the cable, calculate the maximum value point of the cross-correlation function of the signals at both ends, and obtain the accurate time difference. S44. Combine the cable length parameter, corrected propagation speed, and accurate time difference to calculate the position of the discharge source.

2. The method for identifying and locating partial discharge of high-voltage cables according to claim 1, characterized in that, Step S2 includes: S21. Select the wavelet basis function for the collected partial discharge pulse signal. S22. Perform n-layer wavelet decomposition on the partial discharge pulse signal to obtain the approximation coefficient a n and the detail coefficients d1 - d n ; S23. Calculate the signal-to-noise ratio SNR of the detail coefficients of each layer j :[[]]END]] ； Wherein, is the signal variance, is the noise variance, and j = 1, 2, ..., n represents the number of decomposition levels; S24. Establish an adaptive weight coefficient calculation model: ； where w j is the adaptive weight coefficient; S25. Based on the adaptive weight coefficient and the detail coefficients of the corresponding layer, perform wavelet reconstruction according to the reconstructed signal model to obtain the denoised partial discharge feature signal, where the reconstructed signal model is: ； where Y(t) represents the reconstructed signal, T j (d j ) represents the threshold function.

3. The method for identifying and locating partial discharge of high-voltage cable according to claim 2, characterized in that, Threshold function T j (d j ) The soft threshold function is adopted: ； Among them, is the adaptive threshold of the j-th layer: ； Wherein, is the standard deviation of the noise of the j-th layer, and N is the signal length.

4. The method for identifying and locating partial discharge of high-voltage cables according to claim 2, characterized in that, Step S2 also includes: Determine the optimal decomposition scale n based on the signal frequency feature of the partial discharge pulse signal. ； Wherein, is the sampling frequency, is the lowest frequency of interest, and floor is the floor function.

5. The method for identifying and locating partial discharge of high-voltage cables according to claim 1, characterized in that, The correction formula for the propagation speed is: ； wherein, is the corrected propagation speed, is the theoretical propagation speed in the high-voltage cable, is the discharge type correction coefficient, is the discharge intensity correction coefficient, type is the discharge type coefficient, and strength is the discharge intensity coefficient; The formula for the accurate time difference is: ； In the formula, is the precise time difference, is the cross-correlation function of the signals at both ends. , is the signal collected at the first end, is the signal collected at the second end, and t is the time delay value to be tried, means shifting the second-end signal to the right by t time units; The formula for the comprehensive correction term is: ； In the formula, is the comprehensive correction term, freq is the main frequency feature of the signal, amp is the amplitude feature of the signal, shape is the waveform feature of the signal, is the weight coefficient of the corresponding feature; The formula for calculating the position of the discharge source is: ； Wherein, d is the distance from the partial discharge point to the reference end, and L is the total length of the cable.

6. The method for identifying and locating partial discharge of high-voltage cables according to claim 5, characterized in that, The discharge type discrimination result includes three categories, namely internal discharge, surface discharge, and corona discharge. Correspondingly, the values of the discharge type coefficient type are: Internal discharge: type = 1.0; Surface discharge: type = 0.8; Corona discharge: type = 0.6; The formula for calculating the discharge intensity coefficient strength is: strength = min(1.0, q / q0); In the formula, q is the actual discharge intensity, and q0 is the theoretical discharge intensity threshold.

7. The method for identifying and locating partial discharge of high-voltage cables according to claim 1, characterized in that, The sampling accuracy of the high-frequency current transformer is 14 bits, the sampling frequency is 10 - 100 MS / s, and the bandwidth is 0.01 - 10 MHz.

Images