Anti-noise enhanced power distribution network cable fault detection and positioning method

Abstract

The invention provides an anti-noise enhanced power distribution network cable fault detection and positioning method. Multi-period test signals are injected, reflected signals are collected, m-sequence denoising characteristics and self-adaptive wavelet transform are combined, denoising and normalization processing is carried out on the multi-period reflected signals, and the signal quality is improved. The power spectrum functions are averaged after multi-cycle quadratic cross-correlation weighting, and anti-noise enhanced fault positioning is realized through IFFT processing. Meanwhile, filtering signal time domain features (such as amplitude, phase and envelope) of a first period are extracted and fused with a final power spectrum function IFFT value after secondary cross-correlation weighting into a high-dimensional feature vector, and the high-dimensional feature vector is input into a deep learning model based on a convolutional neural network, LSTM and an attention mechanism. The attention mechanism is used for weighting the IFFT value of the power spectrum function after quadratic cross-correlation weighting, extracting peak value characteristics, and improving the positioning precision. The method realizes accurate classification and positioning of cable faults, has high precision and anti-interference capability, is suitable for cable fault detection in a complex environment, and provides technical support for stable operation of a power system.

Description

Technical Field

[0001] The present application relates to the fields of computer technology and signal processing, and in particular to a method for detecting and locating cable faults in a distribution network with enhanced noise resistance. Background Art

[0002] With the expansion of the scale of power systems and the growth of electricity demand, cables are increasingly used in urban distribution networks, especially low-voltage and medium-voltage cables, which are important media for power transmission and bear critical loads. However, the complex laying environment makes cables susceptible to external interference, aging, humidity and other factors, resulting in an increase in the frequency of short-circuit and open-circuit faults year by year. Cable faults not only cause local power outages, but may also have a chain reaction on the overall stability of the power grid. Therefore, it is crucial to quickly and accurately locate cable faults and determine their types to ensure the operation of the power grid. Improving cable fault detection and location technology can effectively reduce maintenance pressure. Rapid location and identification of fault types can help shorten power outages, reduce economic losses and improve power grid reliability. At present, cable fault detection is gradually shifting from traditional physical means to signal processing and intelligent algorithms. Traditional methods such as time domain reflectometers and frequency domain reflectometers are greatly affected by noise in low-voltage cable applications and are difficult to meet the detection needs of complex environments. Therefore, combining intelligent algorithms based on deep learning with signal processing to extract and locate fault features is gradually becoming a new trend in cable fault detection. Summary of the invention

[0003] In view of the above problems, the present invention provides a distribution network cable fault detection and positioning method with enhanced noise resistance, which can realize rapid and accurate positioning and type identification of low-voltage cable faults, so as to improve detection accuracy and anti-interference ability.

[0004] In a first aspect, the present invention adopts a method for detecting and locating cable faults in a distribution network with enhanced noise immunity, comprising the following steps:

[0005] S1: Establish a model of cable fault signal, inject m complete m-sequence periodic test signals into the cable and set it as:

[0006]

[0007] x i (n) is the test signal of the ith complete m-sequence period, y i (n) corresponds to x i(n) The reflection signal of the ith complete cycle, where a is the attenuation coefficient, d is the delay time of the reflection signal, and v(n) is the noise. Multiple complete cycles of the m-sequence modulated test signal are injected into the cable to collect the reflection signal in the cable. Each complete cycle of the reflection signal is windowed and filtered through a bandpass filter to remove the invalid low-frequency band and high-frequency part of the m-sequence modulated test signal, and each complete m-sequence period reflection signal is subjected to adaptive wavelet denoising to obtain the filtered reflection signal y of m complete m-sequence cycles. i ′(n)i=1,2,3...m.

[0008] S2: for the test signal x i (n) and the reflected signal y′ of a complete m sequence period after filtering i (n) Basic cross-correlation operation to obtain a preliminary cross-correlation function Assuming that the noise is ideal Gaussian noise, the result of the cross-correlation calculation can be simplified to A significant peak appears at the delay time τ=d, which provides a reference for the preliminary delay estimation and lays the foundation for the identification of the fault location.

[0009] S3: To further enhance the anti-noise performance, an enhanced secondary cross-correlation operation is performed on the test signal and the reflected signal. First, for each complete cycle of the test signal x i (n) Carry out autocorrelation calculation to obtain R ii (n) Cross-correlation function, extract its autocorrelation characteristics; the quadratic cross-correlation calculation formula is as follows:

[0010]

[0011] Where R ii (n) is the autocorrelation function of the test signal in the ith period, is the cross-correlation function between the ith filtered signal and the ith incident test signal

[0012] When the quadratic cross-correlation operation extracts the enhanced peak characteristics, the quadratic cross-correlation operation is first transformed into a power spectrum function by FFT, and the peak is sharpened by combining weighting processing. The selection of the weighting method needs to consider two factors: noise and reflection. After experimental comparison, it is found that the use of the improved SCOT weighting method makes the peak of the correlation function sharper and more obvious. The improved SCOT weighting method is as follows:

[0013]

[0014] in To improve the weighting function, The test signal x i The power spectrum of (n), for The Fourier transform of is the filtered reflection signal y i The power spectrum of ′(n), is the IFFT processing value of the power spectrum function after the secondary cross-correlation weighting, combined with the anti-noise performance characteristics of the m sequence itself, for multiple periods Averaging:

[0015] Get R′ final (τ) is the final IFFT processing value of the power spectrum function after the secondary cross-correlation weighting, which integrates the anti-noise performance of multiple complete cycles of the m-sequence reflection signal to achieve superposition enhancement of the anti-noise performance. R′ final The time corresponding to the (τ) peak point is the delay time between the two signals. Multiplying it by the propagation speed is the distance of the fault relative to the signal propagation source. While suppressing noise interference, the peak detection accuracy of the signal is further improved. These enhanced characteristics help to improve the recognition accuracy of fault signals.

[0016] S4: The extracted first cycle of the filtered reflection signal y′1(n) time domain features (such as amplitude, phase, envelope) and the final quadratic cross-correlation weighted transform value R′ final (τ) performs multi-dimensional fusion to generate a high-dimensional feature vector. The feature vector is standardized and data is sorted to ensure that the data format is suitable for the input of the subsequent deep learning model. The classification information contained in the time domain features is used to obtain the final quadratic cross-correlation weighted transform value R′ final The positioning information contained in (τ) provides comprehensive feature data for fault classification and fault location.

[0017] S5: Design and build a deep learning model based on convolutional neural network (CNN) and long short-term memory network (LSTM). CNN is used to extract the spatiotemporal features of the signal and identify the local change pattern of the signal; then LSTM is used to capture the timing dependency characteristics of the signal. The sorted high-dimensional feature vector is input into the CNN-LSTM network, which automatically learns and optimizes the characteristic pattern of the signal, outputs the fault type (such as open circuit, short circuit, etc.), and combines the attention mechanism to accurately output the fault location, so as to achieve accurate fault classification and location.

[0018] S6: Output the output of the CNN-LSTM model, i.e., the classification and precise location of the fault, to a display device or storage device for maintenance personnel to view. At the same time, the information can be fed back to the system control center for further maintenance decision-making and scheduling to ensure timely response and safe operation of the power grid.

[0019] Further, the step S1 is specifically as follows: using an m sequence with good autocorrelation characteristics to modulate 2PSK as a test signal input cable, since the available frequency band range of the M sequence is determined by the clock frequency f m The minimum frequency f of the M sequence is constrained by L satisfy The highest frequency f of the M sequence H satisfy So the effective range f L ~f H between.

[0020] Then, according to the effective frequency, design the frequency range f L ~f H The bandpass filter in the filter performs preliminary filtering on the reflected signal and collects the reflected signal y i (n) Normalization is performed to reduce the impact of amplitude differences and enhance the contrast of the signal. The normalization equation is as follows:

[0021]

[0022] Where: x(n) is the signal to be normalized, including the test signal x i (n) and the reflected signal y i (n); is the normalized signal; min(x) and max(x) are the minimum and maximum values ​​of the signal x(n), respectively.

[0023] Further, the step S5 is specifically as follows: in this step, the extracted signal time domain features (such as amplitude, phase, envelope) are multi-dimensionally fused with the final quadratic cross-correlation weighted transform value to generate a high-dimensional feature vector. The feature vector needs to be standardized and the data format is sorted according to the neural network input requirements. The feature extraction and processing steps for the first complete cycle of the filtered reflection signal y1′(n) are as follows:

[0024] Average amplitude A:

[0025]

[0026] Phase θ(n): Instantaneous phase obtained by Fourier transform:

[0027]

[0028] Envelope E(n): The envelope of a signal, usually calculated by Hilbert transform:

[0029]

[0030] in, is the Hilbert transform of the signal y1′(n).

[0031] R′ final (τ): Final quadratic cross-correlation weighted power spectrum function IFFT processing value

[0032] Combine the above features into an initial feature vector: f = [A, θ(n), E(n), R′ final (τ)]

[0033] Each feature in the initial feature vector is standardized to ensure that the numerical ranges of different features are similar to improve the stability of neural network training. The standardization formula is as follows:

[0034]

[0035] Where: f i represents the i-th feature in the feature vector f; is the feature f i The mean of is the feature f i The standard deviation of are the standardized features.

[0036] The standardized features are combined into standardized feature vectors:

[0037]

[0038] The standardized feature vectors are organized into a matrix or tensor format according to the input format of the neural network so that they can be input into the deep learning model. Assuming that the input of the CNN-LSTM neural network is the features of T time steps, the features can be organized into a T×F matrix, where F is the number of features:

[0039]

[0040] Each row represents the features of a time step, and each column represents various types of features. The feature matrix X can be used as the input of a deep learning model (such as CNN-LSTM) to support fault classification and location.

[0041] Furthermore, step S6 is specifically as follows: applying the attention mechanism to the hidden state sequence of LSTM, and focusing on the features that contribute to the fault location prediction by calculating the peak attention weight of each time step. t The calculation formula is

[0042]

[0043] Among them, W a and b aare the parameters of the attention layer, and v is the attention vector.

[0044] The attention weights are combined with the hidden state h t Weighted summation generates weighted feature representation c:

[0045]

[0046] The weighted feature representation c is input into another fully connected layer to predict the specific location of the fault. The output is passed through a linear activation function to obtain the predicted value of the fault location. The formula is:

[0047] The beneficial effects of the present invention are:

[0048] (1) It is proposed to utilize the denoising characteristics of the m-sequence itself, combined with adaptive wavelet transform, to process multi-period reflection signals to achieve a higher denoising effect. And by averaging the multi-period quadratic cross-correlation weighted power spectrum function to obtain the final quadratic cross-correlation weighted power spectrum function, and then through IFFT processing value, a more accurate positioning effect is achieved. The computational complexity of this method is lower than that of other methods. And it can achieve a good denoising effect, so that the present invention can effectively resist some noise in the cable laying environment, realize the stable detection of fault signals, and improve the robustness and reliability of the fault detection system.

[0049] (2) Through multi-dimensional feature fusion input neural network, and the structure of CNN-LSTM network combined with attention mechanism is adopted. CNN is used to extract the spatial features of the signal, LSTM captures the temporal dependency of the signal, and the attention mechanism gives higher weight to the peak features related to the fault location. In this way, even in the environment of low signal-to-noise ratio and strong signal reflection interference, the model can accurately identify the fault type (such as open circuit, short circuit), and more accurately locate the fault location by weighting the peak value of the power spectrum function IFFT processing value after the final quadratic cross-correlation weighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A flowchart of a cable fault detection and location method provided in an embodiment of the present application;

[0052] Figure 2A schematic diagram of the structure of a cable fault detection and locating device provided in an embodiment of the present application.

[0053] Figure 3 A schematic diagram of the structure of a CNN-LSTM model for cable fault detection and location provided in an embodiment of the present application; DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of this application clearer, the relevant technical solutions will be described in detail below with reference to the accompanying drawings. It should be pointed out that the embodiments described here are only part of the content of this application, not all embodiments. In this field, all other embodiments that can be obtained based on the embodiments in this application by ordinary technicians without creative work should be included in the scope of protection of this application.

[0055] Figure 1 A flowchart of a cable fault detection and location method provided in an embodiment of the present application, the improved cable fault detection and location method based on CNN-LSTM specifically includes the following steps:

[0056] Step S101: signal acquisition.

[0057] The design process of the detection system module is shown in Figure 2 shown.

[0058] As an embodiment, the preset test signal is set to an m-sequence with a period N of 511, wherein multiple generation settings are: [1 0 0 0 0 1 0 0 0 1], the initial value is set to: [0 0 0 0 0 00 0 1], the system sampling frequency is set to the clock sampling frequency of 100Mhz, the effective range is 0.1953~33Mhz, and the maximum detection length is set to 100Mhz. The minimum distance is T c is the period of a code element, N is the period length of the signal, Distributed inductance L p And distributed capacitance C p When L p Take 2.45e-5H / km, C p Take 1.2e-6F / km, which can be injected into the cable with a maximum length of 472m and a minimum length of 1m through a capacitive coupler to stimulate the reflection signal that may be caused by the fault. The above only represents one way to set up the test signal. The reflection signal is received through another capacitive coupler and collected using a sensor device, which contains important information about the internal fault of the cable.

[0059] Step S102: signal preprocessing.

[0060] As an embodiment, the collected reflection signal is subjected to denoising, and the signal is passed through an adaptive filter to eliminate high-frequency noise interference and retain the main components of the signal to ensure signal clarity. At the same time, the amplitude of the denoised signal is normalized to facilitate the consistency of feature analysis and processing in subsequent steps.

[0061] The normalized equation is as follows:

[0062]

[0063] Where: x(n) is the signal to be normalized, including the test signal x i (n) and the reflected signal y i (n); is the normalized signal; min(x) and max(x) are the minimum and maximum values ​​of the signal x(n), respectively.

[0064] Step S103: feature extraction.

[0065] After signal preprocessing is completed, the key time domain features and quadratic cross-correlation weighted transform values ​​are extracted from the reflected signal. The time domain features (including the amplitude, phase, and envelope of the signal) can provide the instantaneous change of the signal and reflect the timing information of the fault. First, for each complete cycle of the test signal x i (n) Carry out autocorrelation calculation to obtain R ii (n+τ) cross-correlation function, extract its autocorrelation characteristics; perform secondary cross-correlation operation on the autocorrelation result and the cross-correlation result:

[0066]

[0067] Where R ii (n) is the autocorrelation function of the test signal in the ith period, is the cross-correlation function between the ith filtered signal and the ith incident test signal

[0068] When the quadratic cross-correlation operation extracts the enhanced peak characteristics, the quadratic cross-correlation operation is transformed by FFT to obtain the power spectrum function combined with the quadratic cross-correlation operation, and the power spectrum function of the quadratic cross-correlation operation is weighted to sharpen the peak. The selection of the weighting method needs to consider two factors: noise and reflection. After experimental comparison, it is found that the use of the improved SCOT weighting method makes the peak of the correlation function sharper and more obvious. The improved SCOT weighting method is as follows:

[0069]

[0070] in To improve the weighting function, The test signal xi The power spectrum of (n), for The Fourier transform of is the filtered reflection signal y i The power spectrum of ′(n), is the IFFT processing value of the power spectrum function after the secondary cross-correlation weighting, combined with the anti-noise performance characteristics of the m sequence itself, for multiple periods Averaging:

[0071] Get R′ final (τ) is the final IFFT processing value of the power spectrum function after the secondary cross-correlation weighting, which integrates the anti-noise performance of multiple complete cycles of the m-sequence reflection signal to achieve superposition and enhancement of the anti-noise performance. R′ final The time corresponding to the (τ) peak point is the delay time between the two signals. Multiplying it by the propagation speed is the distance between the fault and the signal propagation source. This suppresses noise interference while further improving the peak detection accuracy of the signal. These enhanced features help to improve the accuracy of fault signal identification.

[0072] Step S104: feature fusion and standardization.

[0073] The extracted multiple time domain features are combined with the final quadratic cross-correlation weighted transform value R′ final (τ) are fused into a high-dimensional feature vector to form a more comprehensive description of signal features. The general algorithm is based only on time domain or frequency domain features, which makes it difficult to simultaneously capture the dynamic changes and enhanced characteristics of the fault signal. Compared with the general algorithm, this algorithm more comprehensively characterizes the time, space and enhanced peak information of the signal, especially in the case of low signal-to-noise ratio.

[0074] These fused feature vectors are then normalized to normalize the different features to a uniform numerical range, and the steps for obtaining the filtered reflection signal y1′(n) of the first complete cycle are:

[0075] Find the average amplitude A of the removed signal:

[0076] Phase θ(n): Instantaneous phase obtained by Fourier transform:

[0077] Envelope E(n): The envelope of a signal, usually calculated by Hilbert transform:

[0078] in, is the Hilbert transform of the signal y1′(n).

[0079] R′ final(τ) is the final quadratic cross-correlation weighted transform value.

[0080] Combine the above features into an initial feature vector: f = [A, θ(n), E(n), R′ final (τ)]

[0081] Each feature in the initial feature vector is standardized to ensure that the numerical ranges of different features are similar to improve the stability of neural network training. The standardization formula is as follows:

[0082]

[0083] Where: f i represents the i-th feature in the feature vector f; is the feature f i The mean of is the feature f i The standard deviation of are the standardized features.

[0084] Step S105, construct a CNN-LSTM model and input data.

[0085] Among them, the schematic diagram of the CNN-LSTM model structure is as follows Figure 3 shown.

[0086] The standardized high-dimensional feature vector is input into the convolutional neural network (CNN), and the spatial features of the signal are extracted by CNN to identify the local characteristic patterns of different faults. These spatial features are then input into the LSTM, and the temporal dependency of the signal is captured by the LSTM temporal memory capability, reflecting the dynamic changes of the signal over time, providing a basis for the fault classification. Finally, the attention mechanism is applied to the R′ output by the LSTM. final The features of (τ) are weighted to enhance the model’s focus on peak features and improve the accuracy of fault location recognition.

[0087] Focus on the features that contribute to the fault location prediction. By the attention weight α t The calculation formula is

[0088]

[0089] Among them, W a and b a are the parameters of the attention layer, and v is the attention vector.

[0090] The attention weights are combined with the hidden state h t Weighted summation generates weighted feature representation c:

[0091]

[0092] Classification output: The weighted feature representation c is input into the fully connected layer to predict the type of fault (such as open circuit, short circuit, etc.). The output passes through the softmax activation function to obtain the probability distribution of each fault type.

[0093] The classification formula is:

[0094] Location prediction: The weighted feature representation c is input into another fully connected layer to predict the specific location of the fault. The output is passed through a linear activation function to obtain the predicted value of the fault location.

[0095] The position prediction formula is:

[0096] Step S106: output the fault classification and location results.

[0097] After the CNN-LSTM network is combined with the attention mechanism, the system can accurately identify the type of fault (such as open circuit, short circuit, etc.) and classify different fault modes. At the same time, combined with the key features after weighted processing, the system can accurately output the location of the fault, providing a basis for the specific location of the fault point in the cable.

[0098] In the loss function and training, classification loss (cross entropy loss) and regression loss (mean square error) are used to jointly optimize the network to ensure the accuracy of classification and positioning.

[0099] Classification loss:

[0100]

[0101] Among them, y i is the actual fault type label, is the model's predicted probability for the failure type.

[0102] Position loss:

[0103]

[0104] Among them, d i is the actual fault location (such as the coordinates of the fault point), is the fault location predicted by the model.

[0105] Adopting the independent optimization strategy, for classification tasks and regression tasks, the two loss functions are optimized independently, and the parameters of the model are updated separately during the training process. Specifically, the cross entropy loss (L class ) is used to optimize the classification task and adjust the weights of the model through the back propagation algorithm to maximize the accuracy of the classification. pos) to optimize the position regression task and independently adjust the network weights to minimize the position prediction error. This enables the model to accurately classify the fault type and accurately predict the fault location. This result will provide reference information for timely maintenance and repair of cables, improving detection efficiency and accuracy.

[0106] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting and locating cable faults in a distribution network with enhanced noise immunity, comprising the following steps: Inject m complete m-sequence periodic test signals and collect the reflected signals of the cable; The reflected signals of multiple complete m-sequence cycles collected are denoised. The invalid low-frequency components and high-frequency interference are first eliminated through a bandpass filter and the main components of the signal are retained. The denoising characteristics of the m-sequence itself are combined with adaptive wavelet transform to denoise the multi-cycle reflected signals and obtain the multi-cycle denoised reflected signals. The denoised reflected signal is normalized to normalize the signal amplitude to a preset range to obtain a multi-cycle normalized denoised reflected signal; A secondary cross-correlation calculation is performed on the multi-period normalized denoised reflection signal and the original multi-period test signal, and the secondary cross-correlation calculation value is transformed by FFT to obtain the power spectrum function, and the peak is highlighted by the multiplication weighted function. The weighted power spectrum functions of multiple periods are averaged to obtain the final secondary cross-correlation weighted power spectrum function, and then the final secondary cross-correlation weighted power spectrum function IFFT processing value is obtained after IFFT processing. Extract time domain features from the normalized signal, perform feature fusion with the final quadratic cross-correlation weighted power spectrum function IFFT processing value, and generate a high-dimensional feature vector; The high-dimensional feature vector is input into a deep learning model based on convolutional neural network (CNN), long short-term memory network (LSTM) and attention mechanism. The spatial features of the signal are extracted by CNN, and the temporal dependency characteristics of the signal are captured by LSTM to accurately identify the fault type. The peak value of the final quadratic cross-correlation weighted power spectrum function IFFT processing value is weighted by the attention mechanism to generate weighted features, so as to locate the fault position more accurately and achieve good denoising effect. The deep learning model outputs the fault type and fault location to complete the classification and location of the cable fault.

2. The cable fault detection and location method according to claim 1, wherein the denoising process in step 2) adopts a bandpass filter denoising method to eliminate noise and retain the main components of the signal. The invalid low-frequency components and high-frequency interference are first eliminated by a bandpass filter and the main components of the signal are retained. The denoising characteristics of the m-sequence itself are utilized in combination with an adaptive wavelet transform to denoise the multi-period reflected signal to obtain a multi-period denoised reflected signal.

3. The cable fault detection and location method according to claim 1, wherein the normalization processing in step 3) is achieved by scaling the amplitude of the denoised signal to within the range of [0,1]. The normalized equation is as follows: in: x(n) is the original signal (it can be the test signal x1(n) or the reflected signal x2(n)); is the normalized signal; min(x) and max(x) are the minimum and maximum values ​​of the signal x(n), respectively.

4. According to the cable fault detection and location method of claim 1, the weighted power spectrum functions of multiple cycles are averaged to obtain the final quadratic cross-correlation weighted power spectrum function, and then the final quadratic cross-correlation weighted power spectrum function IFFT processing value is obtained. Achieve effective noise immunity.

5. The cable fault detection and location method according to claim 1, wherein the time domain features extracted in step 4) include the amplitude, phase and envelope of the signal, and the final power spectrum function IFFT processing value after quadratic cross-correlation weighting.

6. The cable fault detection and location method according to claim 1, wherein the convolutional neural network (CNN) in the deep learning model is a one-dimensional convolutional neural network, which is used to extract the spatial local features of the signal.

7. The cable fault detection and location method according to claim 1, wherein the long short-term memory network (LSTM) in the deep learning model is used to capture the timing dependency characteristics of the signal to generate timing features.

8. The cable fault detection and location method according to claim 1, wherein the attention mechanism of the deep learning model is used to weight the peak features of the final quadratic cross-correlation weighted power spectrum function IFFT processing value generated by LSTM to enhance the focus on key features and improve the positioning accuracy of the fault position.

9. A cable fault detection and location device, comprising: A signal acquisition module, used for injecting a test signal and acquiring a reflected signal of the cable; A preprocessing module, used for denoising and normalizing the collected reflection signal; A feature extraction module is used to extract time domain features and the final IFFT processing value of the power spectrum function after quadratic cross-correlation weighting from the preprocessed signal, and generate a high-dimensional feature vector; Deep learning processing module, including convolutional neural network (CNN), long short-term memory network (LSTM) and attention mechanism, is used to extract spatial features, temporal features and weight key peak features from high-dimensional feature vectors; Output module, used to output the type and location of cable faults.

10. The device according to claim 8, wherein the output module outputs the fault type through a Softmax function based on the feature vector of the deep learning processing module, and outputs the fault location through a fully connected layer.

Images