Fault positioning method and system for power cable

Abstract

The invention belongs to the technical field of fault positioning, and particularly relates to a power cable fault positioning method and system, and the method comprises the steps: obtaining real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; through Haar wavelet and Symlets wavelet analysis, gathering the signal mutation points into a signal mutation point feature set; constructing a threshold intelligent optimization model, and calculating to obtain a current pulse wave head time difference; wave head time is identified by using an RNN-CNN architecture, and a unit distance wave head time transformation value of the whole cable is obtained through fitting; and calculating and correcting a fault distance, and combining the corrected fault distance with the position information of the initial end of the cable, thereby determining the position of a fault point. According to the invention, through deep combination of multi-stage wavelet analysis, the threshold intelligent optimization model and the RNN-CNN architecture, rapid and accurate positioning of the power cable fault is realized, and the fault diagnosis efficiency and accuracy are significantly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault location, and in particular relates to a method and system for locating a fault of a power cable. Background Art

[0002] Power cables play a vital role in power transmission systems. However, due to various factors during long-term operation, such as aging, external force damage, insulation moisture, etc., power cables are prone to failure. Accurately and quickly locating the fault point of the power cable is extremely critical to ensure the stable operation of the power system, reduce power outage time and maintenance costs.

[0003] Traditional methods for locating power cable faults mainly use the impedance method and the traveling wave method. However, the impedance method usually determines the fault location based on the impedance change before and after the cable fault point. The impedance method has low positioning accuracy and is easily affected by changes in line parameters. The traveling wave method often finds it difficult to accurately identify the arrival time of the traveling wave head, which will result in large positioning errors. Therefore, it is difficult to meet the strict requirements of modern power cables for rapid and accurate fault location. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for locating power cable faults. Through the deep combination of multi-stage wavelet analysis, threshold intelligent optimization model and RNN-CNN architecture, fast and accurate positioning of power cable faults is achieved, and the fault diagnosis efficiency and accuracy are significantly improved, especially in complex cable networks.

[0005] The present invention provides the following technical solution: a method for locating a fault of a power cable, comprising the following steps: S1: Acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; S2: The real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequencies, and the sub-signals of different frequencies are analyzed by Symlets wavelet to locate the signal mutation points, and the signal mutation points are summarized as a signal mutation point feature set; S3: construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; S4: Divide the cable into sections and set up test points in each section. Inject multi-feature pulse signals into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time. Combined with the signal mutation point feature set, the propagation time change per unit distance of each test point is accurately calculated, and the wave head time change per unit distance of the entire cable section is obtained by fitting. S5: The fault distance is calculated by the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and the fault distance is corrected by dynamic fuzzy logic. The corrected fault distance is combined with the position information of the starting end of the cable to determine the location of the fault point.

[0006] Preferably, in S1, real-time cable current pulse waveform data and fault current pulse waveform data are obtained, and the acquisition process is as follows: Set the sampling frequency to The continuous current signal is sampled discretely, where: ; In the formula, Indicates the highest frequency of the original signal twice as much.

[0007] Preferably, in S2, the real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequency segments, and the decomposition process is as follows: In the case of real-time discrete acquisition of cable current pulse waveform data, the given discrete current pulse waveform data sequence is set to x , where n=1, 2, 3, ..., N, n represents the sample point number of the discrete time series, N is the number of data points, for scale j, j=1, 2, 3, ..., J, and , J is the maximum number of decomposition levels, It indicates the upper limit of the total number of sub-bands obtained by decomposing the signal at the maximum decomposition level J; Based on the scale j and the translation parameter k, the corresponding discrete wavelet function and scale function Defined as: ; ; In the formula, represents the scaling factor, represents the scaling factor, and Respectively represent the wavelet function and scale function after scale and translation transformation; Set the real-time cable current pulse waveform data to ; Calculate the approximate coefficients at scale j , calculated as: ; Calculate the detail coefficient at scale j , calculated as: ; Under discrete data, for scale j and translation parameter k, the corresponding discrete Symlets wavelet function is Defined as: ; In the formula, Represents the discrete Symlets wavelet function of the unscaled wavelet function at scale j and translation parameter k; Symlets wavelet is used to analyze sub-signals of different frequency segments, and the detail coefficient is calculated by discrete wavelet transform. The calculation method is: ; In the formula, represents the detail coefficient of the Symlets wavelet at scale j and translation parameter k, represents the sub-signal sequence after Haar wavelet decomposition, Represents the discrete Symlets wavelet function after scaling and translation.

[0008] Preferably, in S2, Symlets wavelet is used to analyze sub-signals of different frequency segments and locate signal mutation points, and the positioning process is as follows: Set detail factor , the total number of which is M; Calculate detail factor The mean and standard deviation , calculated as: ; Where i represents the index used to traverse the detail coefficients. and represents the scale j and translation parameter k corresponding to the i-th detail coefficient, Indicates that in a given and Next, the detail coefficients are obtained after discrete wavelet transform using Symlets wavelet; ; Calculate the threshold T of the detail coefficient as follows: ; In the formula, represents the empirical coefficient, ; For each scale j and translation parameter k, traverse all detail coefficients ; when When , set the adjacent points to and , which is the detail coefficient of Symlets wavelet at scale j and translation parameters k+1, k-1; when and When is the modulus maximum point, and is marked as the signal mutation point; when , no marking is performed; The signal mutation point marking process is: Create a marker sequence with the same length as the original signal data , Initialized to zero, when detected Corresponding position When is a mutation point, , Indicates the position of any detected mutation point, and sets the signal mutation point feature set to fare. , mc represents the total number of signal mutation points, Indicates the position of the mcth mutation point in the signal.

[0009] Preferably, in S3, a threshold intelligent optimization model is constructed, and the construction process is as follows: Set real-time cable current pulse waveform data The collection is ,in , v is the total number of real-time cable current pulse waveform data points, is the current amplitude of the vth real-time cable current pulse waveform data point, is the time corresponding to the vth real-time cable current pulse waveform data point; Set fault current pulse waveform data Collection ,in , g is the total number of fault current pulse waveform data points, is the current amplitude of the g-th fault current pulse waveform data point, is the time corresponding to the g-th fault current pulse waveform data point; The real-time cable current pulse waveform data set B and the fault current pulse waveform data set E are merged into a total sample set S. The merging process is as follows: ; In the formula, Represents the v+gth data point in the total sample set; , Represents data points The current amplitude, Represents data points time; Set the current amplitude threshold to LO; if , then Mark as , used to indicate the wave head; if , then Mark as , used to indicate non-wave head; In the formula, Represents a marker variable, used to mark Whether it is recognized as a wave head; Get a set of labeled samples ; right Normalization is performed, and the processing process is as follows: ; In the formula, and are the minimum and maximum current amplitudes of the entire sample set, represents the current amplitude after normalization; Construct a threshold intelligent optimization model. The construction process is as follows: Set the input feature vector to , the threshold intelligent optimization model is as follows: ; In the formula, Indicates the probability value of sample XE belonging to the wave head, and the value range is between, represents the model parameter, eh represents the natural constant; The threshold intelligent optimization model is trained through the cross entropy cost function to determine the model parameters , the training process is as follows: ; In the formula, represents the value of the cost function, Represents the true category information of each sample, It represents the probability value that the model predicts that the iqth sample belongs to the wave head, and log represents the logarithmic function; Use the gradient descent algorithm to minimize , update the parameters The formula is: ; In the formula, represents the jcth parameter in the model parameters, represents the learning rate, Represents the cost function Parameters The partial derivative of ; In the formula, Represents the feature vector The jcth parameter in ; Set the number of iterations to BH, when the cost function When it iterates to BH times, stop iterating and get the trained model parameters ; Setting model parameters The threshold value is ; Calculating the threshold Accuracy , calculated as: ; In the formula, Indicates that the threshold is The number of true positives when , that is, the number of validation set samples that are actually wave heads and are correctly judged as wave heads, represents the number of true negative examples, that is, the number of validation set samples that are not actually wave heads and are correctly judged as not wave heads. represents the number of false positive examples, that is, the number of validation set samples that are not actually wave heads but are mistakenly judged as wave heads. is the number of false negative examples, that is, the number of validation set samples that are actually wave heads but are mistakenly judged as not wave heads; Calculating the threshold The recall rate , calculated as: ; Calculating the threshold The accuracy , calculated as: ; Calculating the threshold F1 score , calculated as: ; Set the threshold of F1 score to FDC; when ,determination is the optimal threshold ; when , then no judgment is made is the optimal threshold ; And record the time point when the current pulse waveform data changes from non-wave head to wave head, which are recorded as TR1 and TR2 respectively; Calculate the time difference of the current pulse front , calculated as: .

[0010] Preferably, in S4, a multi-feature pulse signal is injected into the test point, and the receiving end uses the RNN-CNN architecture to identify the wave head time, and the identification process is as follows: Set the pulse signal sequence input to the RNN-CNN architecture to , represents the feature vector at time t; Update the hidden state of the RNN layer. The update process is as follows: ; In the formula, represents the hidden state of the RNN layer at time t, represents the weight matrix input to the hidden layer, represents the weight matrix from hidden layer to hidden layer, represents the bias vector of the hidden layer, represents the hyperbolic tangent activation function, represents the hidden state of the RNN layer at time t-1; Set the CNN layer to use kf convolution kernels, and the convolution kernel is , mf is the size in the time dimension, nf is the size in the feature dimension, for the ieth convolution kernel, the convolution operation formula is: ; In the formula, It represents the eigenvalue obtained at time t after the convolution operation of the ieth convolution kernel. Indicates that the ieth convolution kernel is at position The weight of represents the bias vector of the ieth convolution kernel, Represents the RNN layer at position at time t The hidden state of After the convolution operation, a series of feature maps are obtained , represents the kj-th feature map; Then perform the pooling operation, set the pooling window size, and the feature map after pooling , represents the km-th feature map, where: ; In the formula, Indicates that the km-th feature map is at position after pooling operation The characteristic value of Indicates that the kj-th feature map is pooled at position The characteristic value of Represents the area corresponding to the ja-th pooling window; The pooled feature map is flattened to obtain a one-dimensional vector FBB, which is then mapped through a fully connected layer. Assuming that the fully connected layer has LBB neurons, the calculation formula is: ; In the formula, represents the output of the jbth neuron in the fully connected layer after the activation function. represents the activation function, represents the weight of the fully connected layer, NF represents the length of the one-dimensional vector FBB, represents the bias vector, represents the lbth neuron in the one-dimensional vector FBB; An output layer is used to determine the position information of the wave head time, and the output is set to yhz. The calculation method is: ; In the formula, represents the weight of the output layer, Represents the bias vector of the output layer; Set the range of XVB from 0 to t, and yhz falls into In the example, the wave head time at the iyecth test point is calculated by linear mapping. , the calculation process is: ; Where MLE represents the maximum value of yhz and is used for normalization.

[0011] Preferably, in S4, the propagation time change value per unit distance of each test point is accurately calculated in combination with the mutation point feature set, and the wave head time change value per unit distance of the entire cable is obtained by fitting, and the fitting process is as follows: Calculate the change in propagation time per unit distance at each test point , calculated as: ; In the formula, Represents the distance between the iyec-1th test point and the iyecth test point, Indicates the wave head time at the iyec-1th test point; The fitting result is the wave head time change value per unit distance of the whole cable section. , the fitting process is as follows: ; Where ikkq represents the index variable and nw represents the total number of test points.

[0012] Preferably, in S5, the fault distance is calculated by the current pulse wave head time difference and the wave head time change value per unit distance of the entire cable section, and the calculation method is: ; Where, LJNT represents the fault distance; The fault distance is corrected by dynamic fuzzy logic. The correction process is as follows: ; In the formula, represents the revised fault distance, , and are the membership functions of cable temperature, humidity and aging degree, respectively. , and Corresponding to , and The weight coefficient of .

[0013] Preferably, in S5, the corrected fault distance is combined with the cable starting end position information to determine the location of the fault point, and the determination process is as follows: Set the cable start position coordinates to , calculate the location coordinates of the fault point , calculated as: ; Set the actual laying length of the cable as LCDP; if , determine the coordinates of the fault point precise; if , determine the coordinates of the fault point If it is inaccurate, recalculate through S1 to S4 until until.

[0014] A power cable fault location system, used to implement the above-mentioned power cable fault location method, comprises: an acquisition module, an analysis module, a construction module, a calculation module and a location module; An acquisition module is used to acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; The analysis module is used to preliminarily decompose the real-time cable current pulse waveform data through Haar wavelet to obtain sub-signals of different frequency segments, use Symlets wavelet to analyze the sub-signals of different frequency segments and locate the signal mutation points, and summarize the signal mutation points into a signal mutation point feature set; A construction module is used to construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; The calculation module is used to divide the cable into sections and set up test points in each section. Multi-feature pulse signals are injected into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time, and combines the signal mutation point feature set to accurately calculate the change value of the propagation time per unit distance of each test point and fit the wave head time change value per unit distance of the entire cable section. The positioning module is used to calculate the fault distance through the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and to correct the fault distance through dynamic fuzzy logic, and combine the corrected fault distance with the position information of the starting end of the cable to determine the location of the fault point.

[0015] Technical effects and advantages of the present invention: The present invention adopts a multi-stage wavelet analysis strategy. First, the real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet, and sub-signals of different frequency bands are quickly separated. Then, these sub-signals are deeply analyzed by Symlets wavelet to accurately locate the signal mutation point. This coarse-first-fine wavelet analysis combination method comprehensively and deeply mines the potential feature information in the current pulse waveform data, significantly enhances the accuracy and integrity of data feature extraction, provides solid, reliable and high-resolution data support for subsequent fault diagnosis, and greatly improves the initial data quality of fault diagnosis.

[0016] The threshold intelligent optimization model constructed by the present invention is highly adaptive. Through deep learning of real-time cable current pulse waveform data and fault current pulse waveform data, the model can accurately adapt to the current cable operation status and data characteristics, thereby accurately measuring the current pulse wave head time. This data-driven and feature-fused threshold determination method effectively avoids the subjectivity and limitations of traditional threshold setting methods, significantly improves the accuracy and reliability of current pulse wave head time determination, and can effectively reduce fault location deviations, especially in long-distance, multi-branch complex cable networks.

[0017] The present invention introduces the RNN-CNN architecture into the cable fault detection process, giving full play to the advantages of recurrent neural networks in processing time series data and convolutional neural networks in extracting features. After injecting multi-feature pulse signals into the cable segment test points, the RNN-CNN architecture can perform in-depth analysis and identification of complex wave head time signals, and combine the mutation point feature set obtained by the previous wavelet analysis to accurately calculate the change value of the propagation time per unit distance of each test point, and fit the wave head time transformation value per unit distance of the entire cable. This signal processing method based on deep learning can quickly process massive data, significantly shorten the fault location time, and improve the accuracy of fault distance calculation.

[0018] The present invention corrects the fault distance through dynamic fuzzy logic and determines the precise location of the fault point in combination with the cable starting end location information. This method comprehensively considers the influence of factors such as cable temperature, humidity and aging degree, and further improves the accuracy and reliability of fault location. Overall, the present invention performs well in improving the efficiency and accuracy of fault location, and is particularly suitable for the fault detection needs of complex cable networks in modern power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of the steps of the method for locating a fault of a power cable of the present invention; Figure 2 This is a block diagram of a fault location system for a power cable according to the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0021] Embodiment 1: like Figure 1 As shown, the method for locating a fault of a power cable comprises the following steps: S1: Acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; S2: The real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequencies, and the sub-signals of different frequencies are analyzed by Symlets wavelet to locate the signal mutation points, and the signal mutation points are summarized as a signal mutation point feature set; S3: construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; S4: Divide the cable into sections and set up test points in each section. Inject multi-feature pulse signals into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time. Combined with the signal mutation point feature set, the propagation time change per unit distance of each test point is accurately calculated, and the wave head time change per unit distance of the entire cable section is obtained by fitting. S5: The fault distance is calculated by the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and the fault distance is corrected by dynamic fuzzy logic. The corrected fault distance is combined with the position information of the starting end of the cable to determine the location of the fault point.

[0022] Haar wavelet is a simple basis function based on piecewise constant function, which has symmetry, compact support and orthogonality. Symlets wavelet is a symmetric Daubechies wavelet basis with better symmetry. RNN-CNN architecture is an architecture that combines recurrent neural network RNN ​​and convolutional neural network CNN.

[0023] In S1, real-time cable current pulse waveform data and fault current pulse waveform data are obtained. The acquisition process is as follows: Set the sampling frequency to The continuous current signal is sampled discretely, where: ; In the formula, Indicates the highest frequency of the original signal twice as much.

[0024] In S2, the real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequency segments. The decomposition process is as follows: In the case of real-time discrete acquisition of cable current pulse waveform data, the given discrete current pulse waveform data sequence is set to x , where n=1, 2, 3, ..., N, n represents the sample point number of the discrete time series, N is the number of data points, for scale j, j=1, 2, 3, ..., J, and , J is the maximum number of decomposition levels, It indicates the upper limit of the total number of sub-bands obtained by decomposing the signal at the maximum decomposition level J; Based on the scale j and the translation parameter k, the corresponding discrete wavelet function and scale function Defined as: ; ; In the formula, represents the scaling factor, represents the scaling factor, and Respectively represent the wavelet function and scale function after scale and translation transformation; Set the real-time cable current pulse waveform data to ; Calculate the approximate coefficients at scale j , calculated as: ; Calculate the detail coefficient at scale j , calculated as: ; Under discrete data, for scale j and translation parameter k, the corresponding discrete Symlets wavelet function is Defined as: ; In the formula, Represents the discrete Symlets wavelet function of the unscaled wavelet function at scale j and translation parameter k; Symlets wavelet is used to analyze sub-signals of different frequency segments, and the detail coefficient is calculated by discrete wavelet transform. The calculation method is: ; In the formula, represents the detail coefficient of the Symlets wavelet at scale j and translation parameter k, represents the sub-signal sequence after Haar wavelet decomposition, Represents the discrete Symlets wavelet function after scaling and translation.

[0025] In S2, Symlets wavelet is used to analyze sub-signals of different frequency segments and locate signal mutation points. The positioning process is as follows: Set detail factor , the total number of which is M; Calculate detail factor The mean and standard deviation , calculated as: ; Where i represents the index used to traverse the detail coefficients. and represents the scale j and translation parameter k corresponding to the i-th detail coefficient, Indicates that in a given and Next, the detail coefficients are obtained after discrete wavelet transform using Symlets wavelet; ; Calculate the threshold T of the detail coefficient as follows: ; In the formula, represents the empirical coefficient, ; For each scale j and translation parameter k, traverse all detail coefficients ; when When , set the adjacent points to and , which is the detail coefficient of Symlets wavelet at scale j and translation parameters k+1, k-1; when and When is the modulus maximum point, and is marked as the signal mutation point; when , no marking is performed; The signal mutation point marking process is: Create a marker sequence with the same length as the original signal data , Initialized to zero, when detected Corresponding position When is a mutation point, , Indicates the position of any detected mutation point, and sets the signal mutation point feature set to fare. , mc represents the total number of signal mutation points, Indicates the position of the mcth mutation point in the signal.

[0026] This embodiment adopts a multi-stage wavelet analysis strategy. First, the Haar wavelet is used to perform a preliminary decomposition of the real-time cable current pulse waveform data, which can quickly separate the sub-signals of different frequency segments and preliminarily outline the approximate frequency characteristic framework of the data. On this basis, the Symlets wavelet is further used to analyze the sub-signals of each frequency. Relying on the excellent performance of the Symlets wavelet in time-frequency localization characteristics, the signal mutation point is accurately located. This coarse-first-fine wavelet analysis combination method comprehensively and deeply mines the potential feature information in the current pulse waveform data, significantly enhancing the accuracy and completeness of data feature extraction. Compared with the traditional single wavelet analysis method, it can more effectively capture weak signal changes and complex waveform details, providing solid, reliable and high-resolution data support for subsequent fault diagnosis links, greatly improving the initial data quality of fault diagnosis, thereby ensuring the accuracy and stability of the entire fault detection process from the source.

[0027] In S3, a threshold intelligent optimization model is constructed. The construction process is as follows: Set real-time cable current pulse waveform data The collection is ,in , v is the total number of real-time cable current pulse waveform data points, is the current amplitude of the vth real-time cable current pulse waveform data point, is the time corresponding to the vth real-time cable current pulse waveform data point; Set fault current pulse waveform data Collection ,in , g is the total number of fault current pulse waveform data points, is the current amplitude of the g-th fault current pulse waveform data point, is the time corresponding to the g-th fault current pulse waveform data point; The real-time cable current pulse waveform data set B and the fault current pulse waveform data set E are merged into a total sample set S. The merging process is as follows: ; In the formula, Represents the v+gth data point in the total sample set; , Represents data points The current amplitude, Represents data points time; Set the current amplitude threshold to LO; if , then Mark as , used to indicate the wave head; if , then Mark as , used to indicate non-wave head; In the formula, Represents a marker variable, used to mark Whether it is recognized as a wave head; Get a set of labeled samples ; right Normalization is performed, and the processing process is as follows: ; In the formula, and are the minimum and maximum current amplitudes of the entire sample set, represents the current amplitude after normalization; Construct a threshold intelligent optimization model. The construction process is as follows: Set the input feature vector to , the threshold intelligent optimization model is as follows: ; In the formula, Indicates the probability value of sample XE belonging to the wave head, and the value range is between, represents the model parameter, eh represents the natural constant; The threshold intelligent optimization model is trained through the cross entropy cost function to determine the model parameters , the training process is as follows: ; In the formula, represents the value of the cost function, Represents the true category information of each sample, It represents the probability value that the model predicts that the iqth sample belongs to the wave head, and log represents the logarithmic function; Use the gradient descent algorithm to minimize , update the parameters The formula is: ; In the formula, represents the jcth parameter in the model parameters, represents the learning rate, Represents the cost function Parameters The partial derivative of ; In the formula, Represents the feature vector The jcth parameter in ; Set the number of iterations to BH, when the cost function When it iterates to BH times, stop iterating and get the trained model parameters ; Setting model parameters The threshold value is ; Calculating the threshold Accuracy , calculated as: ; In the formula, Indicates that the threshold is The number of true positives when , that is, the number of validation set samples that are actually wave heads and are correctly judged as wave heads, represents the number of true negative examples, that is, the number of validation set samples that are not actually wave heads and are correctly judged as not wave heads. represents the number of false positive examples, that is, the number of validation set samples that are not actually wave heads but are mistakenly judged as wave heads. is the number of false negative examples, that is, the number of validation set samples that are actually wave heads but are mistakenly judged as not wave heads; Calculating the threshold The recall rate , calculated as: ; Calculating the threshold The accuracy , calculated as: ; Calculating the threshold F1 score , calculated as: ; Set the threshold of F1 score to FDC; when ,determination is the optimal threshold ; when , then no judgment is made is the optimal threshold ; And record the time point when the current pulse waveform data changes from non-wave head to wave head, which are recorded as TR1 and TR2 respectively; Calculate the time difference of the current pulse front , calculated as: .

[0028] The threshold intelligent optimization model constructed in this embodiment is highly innovative and adaptive. It uses rich real-time cable current pulse waveform data and fault current pulse waveform data as samples for in-depth training, so that the model can fully learn the inherent laws and characteristic differences of the current pulse waveform under different working conditions. In the actual operation process, the model works closely with the signal mutation point feature set obtained by wavelet analysis, and determines the optimal threshold through a complex intelligent algorithm based on the multi-dimensional information contained in the feature set. This optimal threshold can accurately adapt to the current cable operation status and data characteristics, and then accurately determine the current pulse wave head time TR1 and TR2 of the monitoring equipment at both ends. Through this data-driven and feature-fused threshold determination method, not only the subjectivity and limitations of the traditional threshold setting method are effectively avoided, but also the accuracy and reliability of the current pulse wave head time measurement are greatly improved. In a long-distance, multi-branch complex cable network, the fault location deviation caused by the wave head time measurement error can be significantly reduced, which effectively improves the stability and accuracy of the entire fault location in a complex environment.

[0029] In S4, a multi-feature pulse signal is injected into the test point, and the receiving end uses the RNN-CNN architecture to identify the wave head time. The identification process is as follows: Set the pulse signal sequence input to the RNN-CNN architecture to , represents the feature vector at time t; Update the hidden state of the RNN layer. The update process is as follows: ; In the formula, represents the hidden state of the RNN layer at time t, represents the weight matrix input to the hidden layer, represents the weight matrix from hidden layer to hidden layer, represents the bias vector of the hidden layer, represents the hyperbolic tangent activation function, represents the hidden state of the RNN layer at time t-1; Set the CNN layer to use kf convolution kernels, and the convolution kernel is , mf is the size in the time dimension, nf is the size in the feature dimension, for the ieth convolution kernel, the convolution operation formula is: ; In the formula, It represents the eigenvalue obtained at time t after the convolution operation of the ieth convolution kernel. Indicates that the ieth convolution kernel is at position The weight of represents the bias vector of the ieth convolution kernel, Represents the RNN layer at position at time t The hidden state of After the convolution operation, a series of feature maps are obtained , represents the kj-th feature map; Then perform the pooling operation and set the pooling window size (set according to your needs). The feature map after pooling , represents the km-th feature map, where: ; In the formula, Indicates that the km-th feature map is at position after pooling operation The characteristic value of Indicates that the kj-th feature map is pooled at position The characteristic value of Represents the area corresponding to the ja-th pooling window; The pooled feature map is flattened to obtain a one-dimensional vector FBB, which is then mapped through a fully connected layer. Assuming that the fully connected layer has LBB neurons, the calculation formula is: ; In the formula, represents the output of the jbth neuron in the fully connected layer after the activation function. represents the activation function, represents the weight of the fully connected layer, NF represents the length of the one-dimensional vector FBB, represents the bias vector, represents the lbth neuron in the one-dimensional vector FBB; An output layer is used to determine the position information of the wave head time, and the output is set to yhz. The calculation method is: ; In the formula, represents the weight of the output layer, Represents the bias vector of the output layer; Set the range of XVB from 0 to t, and yhz falls into In the example, the wave head time at the iyecth test point is calculated by linear mapping. , the calculation process is: ; Where MLE represents the maximum value of yhz and is used for normalization.

[0030] In S4, the propagation time change per unit distance of each test point is accurately calculated by combining the mutation point feature set, and the wave head time change per unit distance of the entire cable is obtained by fitting. The fitting process is as follows: Calculate the change in propagation time per unit distance at each test point , calculated as: ; In the formula, Represents the distance between the iyec-1th test point and the iyecth test point, Indicates the wave head time at the iyec-1th test point; The fitting result is the wave head time change value per unit distance of the whole cable section. , the fitting process is as follows: ; Where ikkq represents the index variable and nw represents the total number of test points.

[0031] In S5, the fault distance is calculated by the current pulse wave head time difference and the wave head time change value per unit distance of the entire cable section. The calculation method is: ; Where, LJNT represents the fault distance; The fault distance is corrected by dynamic fuzzy logic. The correction process is as follows: ; Where, represents the revised fault distance, , and are the membership functions of cable temperature, humidity and aging degree, respectively. , and Corresponding to , and The weight coefficient of .

[0032] In S5, the corrected fault distance is combined with the cable start end position information to determine the location of the fault point. The determination process is as follows: Set the cable start position coordinates to , calculate the location coordinates of the fault point , calculated as: ; Set the actual laying length of the cable as LCDP; if , determine the coordinates of the fault point precise; if , determine the coordinates of the fault point If it is inaccurate, recalculate through S1 to S4 until until.

[0033] This embodiment adopts a gradient descent-based optimization algorithm to reversely adjust relevant parameters in the threshold intelligent optimization model, so as to achieve adaptive dynamic matching between the threshold and the actual operating state of the cable, thereby improving the accuracy and stability of subsequent fault diagnosis.

[0034] At the same time, this embodiment cleverly introduces the RNN-CNN architecture into the cable fault detection process, giving full play to the unique advantages of recurrent neural networks in processing time series data and the powerful capabilities of convolutional neural networks in the field of image feature extraction and pattern recognition. After injecting multi-feature pulse signals into the cable segment test points, the RNN-CNN architecture at the receiving end can perform in-depth analysis and identification of complex wave head time signals, and combine the mutation point feature set obtained by the previous wavelet analysis. Through rigorous mathematical calculations and data fitting, the unit distance propagation time change value of each test point is accurately calculated, and the unit distance wave head time transformation value of the entire cable section is further fitted; This signal processing and analysis method based on deep learning architecture can automatically learn and adapt to various complex signal patterns and change laws in cable networks, effectively overcoming the shortcomings of traditional signal processing methods when facing large-scale, highly complex data. In terms of improving fault location efficiency, the parallel computing capability and efficient feature extraction mechanism of the RNN-CNN architecture can quickly process massive test point data and significantly shorten the time required for fault location. In terms of accuracy, its powerful learning and generalization capabilities can accurately identify slight wave head time changes, thereby ensuring high accuracy of fault distance calculation, providing a strong technical guarantee for the rapid and accurate positioning of cable network faults, especially suitable for the increasingly complex cable network structure and working environment in modern power systems, so as to achieve the effect of rapid and accurate positioning of cable faults.

[0035] In this embodiment, when calculating each formula, each parameter may be dimensionless or dimensionless as needed.

[0036] The verification process is: First, set the cable parameters. The total cable length is 10 km. The preset fault point is at 3.5 km (simulating a short circuit fault). For the test point segmentation, set a test point every 1 km (a total of 10 test points).

[0037] Set the sampling frequency to 40kHz ,satisfy ,Real-time current pulse waveform data (normal state) and fault current pulse waveform data are generated by electromagnetic transient simulation software (such as PSCAD).

[0038] Considering environmental factors, the temperature is set to 25°C, the humidity is set to 60%, the aging degree is set to 5%, and the membership function weight is set to , and They are 0.2, 0.1 and 0.3 respectively.

[0039] Wavelet decomposition and mutation point detection were performed. In the Haar wavelet decomposition, the decomposition layer number J=5, and 5 frequency segment sub-signals (low-frequency approximate signal + high-frequency detail signal) were obtained. The sub-signal sequence distribution was 85% low-frequency segment and 15% high-frequency segment. In the Symlets wavelet analysis, the total number of mutation points mc=12 was detected, distributed near the fault point (3.3-3.6 km interval). Perform intelligent threshold optimization and wave head time determination, merge real-time and fault data for a total of 5,000 samples, divide them into training set (80%) and verification set (20%), and calculate the optimal threshold (The corresponding F1 score is 0.92), and the wave head time measurement results are , ,get .

[0040] The propagation time of the test points is calculated. The RNN-CNN recognition result is that the model accuracy is 98.5% (test set), and the wave head time recognition error is less than 0.5%. Among the wave head times of each test point, for the 1st, 2nd, 6th, and 7th test points, their wave head times are They are 2.1, 4.3, 12.5 and 14.8μs respectively.

[0041] Fitting the propagation time per unit distance yields .

[0042] Calculate and correct the fault distance, initial fault distance , dynamic blur correction, membership value , and are 0.3, 0.2, and 0.1 respectively, then , because the simulation data presets the fault point as 3.5km, the error is 0.3km. When it is 0, the coordinates of the fault point , 3.47km is less than the total length of 10km, so the fault point location is determined to be accurate.

[0043] Embodiment 2: like Figure 2 As shown, the fault location system of the power cable is used to implement the method in Example 1, including: an acquisition module, an analysis module, a construction module, a calculation module and a location module; An acquisition module is used to acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; The analysis module is used to preliminarily decompose the real-time cable current pulse waveform data through Haar wavelet to obtain sub-signals of different frequency segments, use Symlets wavelet to analyze the sub-signals of different frequency segments and locate the signal mutation points, and summarize the signal mutation points into a signal mutation point feature set; A construction module is used to construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; The calculation module is used to divide the cable into sections and set up test points in each section. Multi-feature pulse signals are injected into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time, and combines the signal mutation point feature set to accurately calculate the change value of the propagation time per unit distance of each test point and fit the wave head time change value per unit distance of the entire cable section. The positioning module is used to calculate the fault distance through the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and to correct the fault distance through dynamic fuzzy logic, and combine the corrected fault distance with the position information of the starting end of the cable to determine the location of the fault point.

[0044] Embodiment 3: A fault location device for a power cable comprises: a memory on which a computer processing program is stored; and a processor for executing the computer processing program in the memory to implement the system in Embodiment 2.

Claims

1. A method for locating a fault of a power cable, characterized in that: The following steps are involved: S1: Acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; S2: The real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequencies, and the sub-signals of different frequencies are analyzed by Symlets wavelet to locate the signal mutation points, and the signal mutation points are summarized as a signal mutation point feature set; S3: construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; S4: Divide the cable into sections and set up test points in each section. Inject multi-feature pulse signals into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time. Combined with the signal mutation point feature set, the propagation time change per unit distance of each test point is accurately calculated, and the wave head time change per unit distance of the entire cable section is obtained by fitting. S5: The fault distance is calculated by the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and the fault distance is corrected by dynamic fuzzy logic. The corrected fault distance is combined with the position information of the starting end of the cable to determine the location of the fault point.

2. The method for locating a fault of a power cable according to claim 1, characterized in that: In S1, real-time cable current pulse waveform data and fault current pulse waveform data are obtained, and the acquisition process is as follows: Set the sampling frequency to The continuous current signal is sampled discretely, where: ； In the formula, Indicates the highest frequency of the original signal twice as much.

3. The method for locating a fault of a power cable according to claim 1, characterized in that: In S2, the real-time cable current pulse waveform data is preliminarily decomposed by Haar wavelet to obtain sub-signals of different frequency segments. The decomposition process is as follows: In the case of real-time discrete acquisition of cable current pulse waveform data, the given discrete current pulse waveform data sequence is set to x , where n=1, 2, 3, ..., N, n represents the sample point number of the discrete time series, N is the number of data points, for scale j, j=1, 2, 3, ..., J, and , J is the maximum number of decomposition levels, It indicates the upper limit of the total number of sub-bands obtained by decomposing the signal at the maximum decomposition level J; Based on the scale j and the translation parameter k, the corresponding discrete wavelet function and scale function Defined as: ； ； In the formula, represents the scaling factor, represents the scaling factor, and Respectively represent the wavelet function and scale function after scale and translation transformation; Set the real-time cable current pulse waveform data to ; Calculate the approximate coefficients at scale j , calculated as: ； Calculate the detail coefficient at scale j , calculated as: ； Under discrete data, for scale j and translation parameter k, the corresponding discrete Symlets wavelet function is Defined as: ； In the formula, Represents the discrete Symlets wavelet function of the unscaled wavelet function at scale j and translation parameter k; Symlets wavelet is used to analyze sub-signals of different frequency segments, and the detail coefficient is calculated by discrete wavelet transform. The calculation method is: ； In the formula, represents the detail coefficient of the Symlets wavelet at scale j and translation parameter k, represents the sub-signal sequence after Haar wavelet decomposition, Represents the discrete Symlets wavelet function after scaling and translation.

4. The method for locating a fault of a power cable according to claim 3, characterized in that: In S2, Symlets wavelet is used to analyze sub-signals of different frequency segments and locate signal mutation points. The positioning process is as follows: Set detail factor , the total number of which is M; Calculate detail factor The mean and standard deviation , calculated as: ； Where i represents the index used to traverse the detail coefficients. and represents the scale j and translation parameter k corresponding to the i-th detail coefficient, Indicates that in a given and Next, the detail coefficients are obtained after discrete wavelet transform using Symlets wavelet; ； Calculate the threshold T of the detail coefficient as follows: ； In the formula, represents the empirical coefficient, ; For each scale j and translation parameter k, traverse all detail coefficients ; when When , set the adjacent points to and , which is the detail coefficient of Symlets wavelet at scale j and translation parameters k+1, k-1; when and When is the modulus maximum point, and is marked as the signal mutation point; when , no marking is performed; The signal mutation point marking process is: Create a marker sequence with the same length as the original signal data , Initialized to zero, when detected Corresponding position When is a mutation point, , Indicates the position of any detected mutation point, and sets the signal mutation point feature set to fare. , mc represents the total number of signal mutation points, Indicates the position of the mcth mutation point in the signal.

5. The method for locating a fault of a power cable according to claim 1, characterized in that: In S3, a threshold intelligent optimization model is constructed, and the construction process is as follows: Set real-time cable current pulse waveform data The collection is ,in , v is the total number of real-time cable current pulse waveform data points, is the current amplitude of the vth real-time cable current pulse waveform data point, is the time corresponding to the vth real-time cable current pulse waveform data point; Set fault current pulse waveform data Collection ,in , g is the total number of fault current pulse waveform data points, is the current amplitude of the g-th fault current pulse waveform data point, is the time corresponding to the g-th fault current pulse waveform data point; The real-time cable current pulse waveform data set B and the fault current pulse waveform data set E are merged into a total sample set S. The merging process is as follows: ； In the formula, Represents the v+gth data point in the total sample set; , Represents data points The current amplitude, Represents data points time; Set the current amplitude threshold to LO; if , then Mark as , used to indicate the wave head; if , then Mark as , used to indicate non-wave head; In the formula, Represents a marker variable, used to mark Whether it is recognized as a wave head; Get a set of labeled samples ; right Normalization is performed, and the processing process is as follows: ； In the formula, and are the minimum and maximum current amplitudes of the entire sample set, represents the current amplitude after normalization; Construct a threshold intelligent optimization model. The construction process is as follows: Set the input feature vector to , the threshold intelligent optimization model is as follows: ； In the formula, Indicates the probability value of sample XE belonging to the wave head, and the value range is between, represents the model parameter, eh represents the natural constant; The threshold intelligent optimization model is trained through the cross entropy cost function to determine the model parameters , the training process is as follows: ； In the formula, represents the value of the cost function, Represents the true category information of each sample, It represents the probability value that the model predicts that the iqth sample belongs to the wave head, and log represents the logarithmic function; Use the gradient descent algorithm to minimize , update the parameters The formula is: ； In the formula, represents the jcth parameter in the model parameters, represents the learning rate, Represents the cost function Parameters The partial derivative of ； In the formula, Represents the feature vector The jcth parameter in ; Set the number of iterations to BH, when the cost function When it iterates to BH times, stop iterating and get the trained model parameters ; Setting model parameters The threshold value is ; Calculating the threshold Accuracy , calculated as: ； In the formula, Indicates that the threshold is The number of true positives when , that is, the number of validation set samples that are actually wave heads and are correctly judged as wave heads, represents the number of true negative examples, that is, the number of validation set samples that are not actually wave heads and are correctly judged as not wave heads. represents the number of false positive examples, that is, the number of validation set samples that are not actually wave heads but are mistakenly judged as wave heads. is the number of false negative examples, that is, the number of validation set samples that are actually wave heads but are mistakenly judged as not wave heads; Calculating the threshold The recall rate , calculated as: ； Calculating the threshold The accuracy , calculated as: ； Calculating the threshold F1 score , calculated as: ； Set the threshold of F1 score to FDC; when ,determination is the optimal threshold ; when , then no judgment is made is the optimal threshold ; And record the time point when the current pulse waveform data changes from non-wave head to wave head, which are recorded as TR1 and TR2 respectively; Calculate the time difference of the current pulse front , calculated as: 。 6. The method for locating a fault of a power cable according to claim 5, characterized in that: In S4, a multi-feature pulse signal is injected into the test point, and the receiving end uses the RNN-CNN architecture to identify the wave head time. The identification process is as follows: Set the pulse signal sequence input to the RNN-CNN architecture to , represents the feature vector at time t; Update the hidden state of the RNN layer. The update process is as follows: ； In the formula, represents the hidden state of the RNN layer at time t, represents the weight matrix input to the hidden layer, represents the weight matrix from hidden layer to hidden layer, represents the bias vector of the hidden layer, represents the hyperbolic tangent activation function, Represents the hidden state of the RNN layer at time t-1; Set the CNN layer to use kf convolution kernels, and the convolution kernel is , mf is the size in the time dimension, nf is the size in the feature dimension, for the ieth convolution kernel, the convolution operation formula is: ； In the formula, It represents the eigenvalue obtained at time t after the convolution operation of the ieth convolution kernel. Indicates that the ieth convolution kernel is at position The weight of represents the bias vector of the ieth convolution kernel, Represents the RNN layer at position at time t The hidden state of After the convolution operation, a series of feature maps are obtained , represents the kj-th feature map; Then perform the pooling operation, set the pooling window size, and the feature map after pooling , represents the km-th feature map, where: ； In the formula, Indicates that the km-th feature map is at position after pooling operation The characteristic value of Indicates that the kj-th feature map is pooled at position The characteristic value of Represents the area corresponding to the ja-th pooling window; The pooled feature map is flattened to obtain a one-dimensional vector FBB, which is then mapped through a fully connected layer. Assuming that the fully connected layer has LBB neurons, the calculation formula is: ； In the formula, represents the output of the jbth neuron in the fully connected layer after the activation function. represents the activation function, represents the weight of the fully connected layer, NF represents the length of the one-dimensional vector FBB, represents the bias vector, represents the lbth neuron in the one-dimensional vector FBB; An output layer is used to determine the position information of the wave head time, and the output is set to yhz. The calculation method is: ； In the formula, represents the weight of the output layer, Represents the bias vector of the output layer; Set the range of XVB from 0 to t, and yhz falls into In the example, the wave head time at the iyecth test point is calculated by linear mapping. , the calculation process is: ； Where MLE represents the maximum value of yhz and is used for normalization.

7. The method for locating a fault of a power cable according to claim 6, characterized in that: In S4, the propagation time change value per unit distance of each test point is accurately calculated in combination with the mutation point feature set, and the wave head time change value per unit distance of the entire cable is obtained by fitting. The fitting process is as follows: Calculate the change in propagation time per unit distance at each test point , calculated as: ； In the formula, Represents the distance between the iyec-1th test point and the iyecth test point, Indicates the wave head time at the iyec-1th test point; The fitting result is the wave head time change value per unit distance of the whole cable section. , the fitting process is as follows: ； Where ikkq represents the index variable and nw represents the total number of test points.

8. The method for locating a fault of a power cable according to claim 7, characterized in that: In S5, the fault distance is calculated by the current pulse wave head time difference and the wave head time change value per unit distance of the entire cable section, and the calculation method is: ； Where, LJNT represents the fault distance; The fault distance is corrected by dynamic fuzzy logic. The correction process is as follows: ； In the formula, represents the revised fault distance, , and are the membership functions of cable temperature, humidity and aging degree, respectively. , and Corresponding to , and The weight coefficient of .

9. The method for locating a fault of a power cable according to claim 8, characterized in that: In S5, the corrected fault distance is combined with the cable starting end position information to determine the location of the fault point. The determination process is as follows: Set the cable start position coordinates to , calculate the location coordinates of the fault point , calculated as: ； Set the actual laying length of the cable as LCDP; if , determine the coordinates of the fault point precise; if , determine the coordinates of the fault point If it is inaccurate, recalculate through S1 to S4 until until.

10. A power cable fault location system, used to implement the power cable fault location method according to any one of claims 1 to 9, characterized in that: include: Acquisition module, analysis module, construction module, calculation module and positioning module; An acquisition module is used to acquire real-time cable current pulse waveform data, fault current pulse waveform data and historical cable operation data; The analysis module is used to preliminarily decompose the real-time cable current pulse waveform data through Haar wavelet to obtain sub-signals of different frequency segments, use Symlets wavelet to analyze the sub-signals of different frequency segments and locate the signal mutation points, and summarize the signal mutation points into a signal mutation point feature set; A construction module is used to construct a threshold intelligent optimization model, determine the optimal threshold by combining the threshold intelligent optimization model with the signal mutation point feature set, accurately measure the current pulse wave head time TR1 and TR2, and calculate the current pulse wave head time difference; The calculation module is used to divide the cable into sections and set up test points in each section. Multi-feature pulse signals are injected into the test points. The receiving end uses the RNN-CNN architecture to identify the wave head time, and combines the signal mutation point feature set to accurately calculate the change value of the propagation time per unit distance of each test point and fit the wave head time change value per unit distance of the entire cable section. The positioning module is used to calculate the fault distance through the time difference of the current pulse wave head and the wave head time change value per unit distance of the entire cable section, and to correct the fault distance through dynamic fuzzy logic, and combine the corrected fault distance with the position information of the starting end of the cable to determine the location of the fault point.

Images