XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network

Abstract

The invention discloses an XLPE cable insulation fault diagnosis model based on an adaptive CEEMDAN and a multi-scale convolutional neural network, and the model comprises the steps: decomposing an original fault signal into a plurality of equal sub-signals through a sliding window technology, and enabling each sub-signal to serve as an independent sample; then, a training set and a verification set are selected according to the random principle, it is ensured that the data samples of the fault types account for half of the training set and half of the verification set respectively, and the proportion is balanced; training set fault data and verification set fault data are reconstructed based on the adaptive CEEMDAN; inputting the reconstructed training set fault data into the MSCNN for training, and obtaining a related parameter with an optimal training result; and inputting the reconstructed verification set fault data into the trained multi-scale convolutional neural network MSCNN to obtain a fault identification result. The ACEEMDAN-MSCNN provided by the invention is superior to a basic single-scale CNN (Convolutional Neural Network) model and other multi-scale models.

Description

Technical Field

[0001] The invention relates to a cable insulation fault diagnosis technology, and in particular to an XLPE cable insulation fault diagnosis model based on an adaptive CEEMDAN and a multi-scale convolutional neural network. Background Art

[0002] At present, with the continuous increase in my country's investment in ultra-high voltage power grids, the requirements for the stability of power equipment are becoming more and more stringent. As an important part of the power system, the reliable operation of XLPE cables is directly related to the safety and stability of the power grid. However, as time goes by, the cables are susceptible to external factors such as mechanical damage, insulation aging, and abnormal working conditions, resulting in a significant increase in the failure rate, and its stability and reliability are reduced, which brings serious safety hazards to the smooth operation of the transmission network. Therefore, it is urgent to develop real-time fault diagnosis technology for high-voltage cables.

[0003] Fault detection technologies for XLPE cables at home and abroad include DC superposition method, DC component method, dielectric loss method, partial discharge detection method, and infrared temperature measurement method. However, these methods have certain limitations. For example, the DC superposition method is difficult to accurately predict the remaining service life of the cable, the DC component method has insufficient measurement accuracy, the dielectric loss method has a limited application range, the partial discharge detection method has a weak signal, and the infrared temperature measurement method has high equipment maintenance costs. These methods still have room for improvement in the efficiency and accuracy of fault detection. In addition, the occurrence of cable faults is unpredictable, and there are many types of faults with a wide distribution range. Many scholars have conducted in-depth research on cable grounding wire current and its detection methods.

[0004] For example, some people have incorporated the influence of insulation layer leakage current and capacitance current into the grounding wire current calculation model, and proposed a fault diagnosis standard based on the change of grounding wire current value. Some people use unsupervised learning and time domain analysis technology to mine the data of cable sheath circulation, which provides a basis for the determination of circulation. Some people have proposed a new online insulation monitoring technology-two-terminal synchronization method, but this method is still affected by harmonics and frequency fluctuations. Some people use cable sheath current to construct Lissajous figures, and combine them with the core temperature change to improve the accuracy of fault identification, but the detection accuracy of complex line structures still needs to be improved. Faced with the high difficulty of high-voltage cable fault detection and the high cost challenge of real-time monitoring, some people have successfully developed a new diagnostic technology that combines the convolutional neural network of the set empirical mode decomposition and the improved longhorn beard search algorithm. At the same time, some scholars have proposed an innovative XLPE cable fault detection method, which uses a deep residual shrinkage network and incorporates a soft threshold nonlinear conversion layer into the network structure. At the same time, the soft threshold is optimized with the soft attention mechanism, which not only enhances the ability of the deep neural network to extract features from the local discharge signal interfered by noise, but also significantly improves the accuracy of cable fault detection.

[0005] Although existing studies have explored the fault data of cross-linked polyethylene cables, these studies have not deeply explored the multi-layer feature information under the fault state. In view of the changing environment faced by XLPE cables in practical applications, the local discharge signals generated by them are often mixed with many interference noises, coupled with the uneven amount of fault data of different cables and the heterogeneity of data dimensions, resulting in the lack of accuracy of traditional detection methods. In response to these problems, the present invention constructs an XLPE cable insulation fault diagnosis model (ACEEMDAN-MSCNN) based on adaptive CEEMDAN and multi-scale convolutional neural network. The model uses CEEMDAN decomposition to perform multi-scale analysis on each data sequence, and then adaptively selects effective IMF components as feature factors, and finally calculates the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension; then, a multi-scale convolutional neural network is used to realize fault detection and rapid classification, thereby achieving the purpose of intelligent high-voltage cable fault diagnosis. Summary of the invention

[0006] The main purpose of the present invention is to provide an XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network. The model first uses CEEMDAN decomposition to perform multi-scale analysis on each data sequence, then adaptively selects effective IMF components as characteristic factors, and finally calculates the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension; then, a multi-scale convolutional neural network model, MSCNN, is proposed, which takes the reconstructed fault signal as input, does not require any manual features, can automatically extract fault-sensitive features, and perform fault classification.

[0007] The technical solution adopted by the present invention is: an XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network, comprising:

[0008] The sliding window technology is used to decompose the original fault signal into several equal sub-signals, each of which is used as an independent sample. Then, the training set and the validation set are selected according to the random principle to ensure that the data samples of the fault type account for half of the training set and the validation set respectively, and the proportion is balanced.

[0009] Reconstruct the training set fault data and the validation set fault data based on adaptive CEEMDAN;

[0010] The reconstructed training set fault data is input into MSCNN for training to obtain relevant parameters with optimal training results;

[0011] The reconstructed verification set fault data is input into the trained multi-scale convolutional neural network MSCNN to obtain the fault recognition result.

[0012] Furthermore, the reconstruction of the training set fault data and the validation set fault data based on the adaptive CEEMDAN includes:

[0013] CEEMDAN is used to perform multi-scale analysis on each data sequence, and then the effective IMF components are adaptively selected as characteristic factors. Finally, the corresponding energy entropy is calculated as the final eigenvalue to reconstruct all data samples with the same dimension. Represents the i-th data sample Y i The jth feature variable in Y i The dimension of .

[0014] Furthermore, the input of the multi-scale convolutional neural network MSCNN is the reconstructed fault data, and the output is a predicted label indicating the fault condition; the input sample X∈R l It is represented by X = {x1, x2, ..., x l}, where l is the length of the input sample, the output Y is a One-Hot vector, and the fault diagnosis task is defined as using the MSCNN model to obtain the label Y based on the reconstructed fault diagnosis data X. The MSCNN model has four parts: denoising layer, AWMSC block, multi-scale feature fusion layer and classification layer.

[0015] Advantages of the present invention:

[0016] The model of the present invention uses CEEMDAN decomposition to perform multi-scale analysis on each data sequence, then adaptively selects effective IMF components as characteristic factors, and finally calculates the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension; then, a multi-scale convolutional neural network is used to realize fault detection and rapid classification, thereby achieving the purpose of intelligent high-voltage cable fault diagnosis.

[0017] ACEEMDAN-MSCNN outperforms basic single-scale CNN models and other multi-scale models.

[0018] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 It is the framework of the MSCNN model of the present invention;

[0021] Figure 2 It is the adaptive calculation process of the weight vector W of the present invention;

[0022] Figure 3 is the convergence performance of different models of the present invention;

[0023] Figure 4 are the fault identification results of different models of the present invention;

[0024] Figure 5 It is a performance comparison of different models of the present invention in a noisy environment with different signal-to-noise ratios. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] Fault data reconstruction based on adaptive CEEMDAN:

[0027] Due to the complex operating environment of XLPE cables, the local discharge signals on site usually contain a large amount of noise signals, and the number of fault diagnosis data for different cables is different, and the dimensions are also different, which brings great challenges to the fault diagnosis of XLPE cables. To address this problem, the present invention proposes a fault data reconstruction method based on adaptive CEEMDAN. The method first uses CEEMDAN decomposition to perform multi-scale analysis on each data sequence, then adaptively selects effective IMF components as characteristic factors, and finally calculates the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension. here, Represents the i-th data sample Y i The jth feature variable in Y i The dimension of .

[0028] Adaptive selection of IMF components:

[0029] As mentioned above, in order to construct data samples Y with the same dimension i (i∈[1,m]), it is necessary to determine the characteristic factors and calculate the corresponding eigenvalues, which can be further used to obtain the characteristic variables In terms of feature factor selection, although IMF components can be used as feature factors for some traditional anomaly detection models, there is still a considerable problem that fixed parameter values ​​cannot accurately describe the intrinsic characteristics of the original data. To solve this problem, the proposed feature development method can fully analyze the contribution of a single IMF component and the global reconstruction error, and adaptively adjust the number of effective IMF components according to the intrinsic characteristics. The specific selection process is as follows:

[0030] In step 1, the root mean square error (RMSE), correlation coefficient, and energy difference between the original data and the reconstructed data are calculated, and the adjustment coefficient β is designed to appropriately adjust the number of IMF components so as to adaptively select effective IMF components.

[0031] Assume that x and x' represent the original data and the reconstructed data, respectively. The numerical difference between x and x' can be measured by RMSE, which is defined as

[0032]

[0033] The correlation between x and x' can be measured by the correlation coefficient, which is defined as:

[0034]

[0035] In addition, the energy difference between x and x' can be calculated by the following formula:

[0036]

[0037] Where E represents the energy value calculation of the original data or the reconstructed data. According to the above parameters, the final adjustment coefficient β can be defined as follows:

[0038]

[0039] Step 2. Calculate the cumulative variance contribution and cumulative energy ratio of each IMF component and dynamically adjust its threshold parameters. After that, further select the effective IMF components that are less than two threshold parameters. The two threshold parameters of the IMF components can be calculated by the following formula:

[0040]

[0041] Here, λ j is the variance of the jth IMF component, E j is the energy of the jth IMF component.

[0042] Since the corresponding probability distribution and energy distribution of the XLPE cable fault diagnosis data sequence will change dynamically, when the data sequence is subjected to multi-scale analysis, the probability distribution and energy distribution of each IMF component obtained by CEEMDAN decomposition will also change. In order to successfully track this change, the information entropy based on the energy distribution of the IMF component is introduced in this paper, and the energy entropy of the effective IMF component is used as the final eigenvalue.

[0043] Characteristic calculation based on CEEMDAN energy entropy:

[0044] The specific steps of feature calculation based on CEEMDAN energy entropy are as follows:

[0045] Step 1 (Data Preprocessing). As mentioned above, each data sequence is obtained from the raw IIoT sensor data. And form a data sequence set D = {D1, D2, D3, ..., D m}.

[0046] Step 2 (IMF component calculation). Each data series Decompose it through CEEMDAN decomposition and get J IMF components.

[0047] First, assume that the original data x = D i , and use CEEMDAN to decompose x+ε0w v Different experiments were performed. Furthermore, the EMD decomposition in each experiment was run until the first EMD modal component was obtained. From these experiments, the first mean IMF component can be further calculated:

[0048]

[0049] Among them, IMF v1 represents the first IMF component of the vth experiment. In addition, the first unique remainder can be obtained as follows:

[0050] r1=x-imf'1 (7)

[0051] Secondly, according to the above method, further decompose the r of each j (j = 1, 2, ..., J) i +ε i E i (w v ),v=1,2,…,I:

[0052]

[0053] In addition, the jth unique remainder can be obtained by the following formula:

[0054] r j =r j-1 -imf' i ,j∈[2,J] (9)

[0055] Among them, IMF' j is the jth IMF component obtained by CEEMDAN decomposition, E j (.) is the first EMD modal component obtained by EMD decomposition, ε j-1 Is to add noise to solve the imf' j The signal-to-noise ratio adjustment factor, w v is the zero-mean white noise source added in experiment v.

[0056] Finally, repeat the above calculation process until there is no remainder to be decomposed, and obtain all J IMF components IMF = {imf'1, imf'2, ..., imf' j}. In addition, the final remainder can be calculated by the following formula:

[0057]

[0058] To sum up, the original data x can finally be decomposed into:

[0059]

[0060] Step 3. According to the adaptive IMF selection process, the RMSE, correlation coefficient and energy difference between the original data x and the reconstructed data x' reconstructed by the IMF components are calculated, and the cumulative variance contribution and cumulative energy ratio of each IMF component are calculated. Through adaptive IMF selection, effective IMF components can be obtained.

[0061] Step 4. By further calculating the energy of each effective IMF component Construct the corresponding energy vector V E =(E1,E2,…,E f ).

[0062] Step 5. For the energy vector V E , the calculated energy entropy of each effective IMF component It can be regarded as an eigenvalue, and the energy entropy vector V can be obtained H =(H1,H2,…,H f ). The energy entropy of each effective IMF component can be calculated by the following formula:

[0063] H(E j )=-P(E j )log P(E j ) (12)

[0064] Among them, E j Represents the energy value of the jth IMF component is the energy proportion of the jth IMF component in the total energy.

[0065] Step 6. Set up data sample The final reconstructed data sample set Y = {Y1, Y2, ..., Y m}.

[0066] Multi-scale Convolutional Neural Networks:

[0067] The proposed multi-scale convolutional neural network, MSCNN, is implemented under the assumption that labeled training data sets under at least two operating conditions are available for model training. The key to the effectiveness of the model is that it can extract multi-scale features and can adaptively modulate the multi-scale features using an adaptive weight vector. The extracted rich features are only sensitive to faults and not to operating conditions.

[0068] The proposed MSCNN framework is as follows Figure 1 As shown in Figure 2, the input of MSCNN is the reconstructed fault data, and the output is the predicted label indicating the fault condition. l It is represented by X = {x1, x2, ..., x l}, where l is the length of the input sample. The output Y is a One-Hot vector. The fault diagnosis task is defined as using the MSCNN model to obtain the label Y based on the reconstructed fault diagnosis data X. The MSCNN model has four parts: denoising layer, AWMSC block, multi-scale feature fusion layer and classification layer.

[0069] Denoising layer

[0070] Before extracting multi-scale features, a denoising layer is applied to suppress background noise. The denoising layer is actually a 1D convolutional (Conv1D) layer.

[0071] The Conv1D layer usually contains multiple 1D convolution kernel channels, each channel is used to extract a feature. Assuming the number of channels is c, the convolution operation can be expressed as:

[0072]

[0073] where F i d , (i=1,2,…,c) is the i-th channel feature in the d-th layer feature map, K i is the convolution kernel of the i-th channel, X d-1 is the feature map of the (d-1)th layer, b i is the bias, and F is the ReLU activation function. c channels can output the feature map F, which is expressed as:

[0074]

[0075] In MSCNN, the parameters of the denoising layer are c1@k1. The width k1 of the 1-dimensional kernel is wider than the kernels in other layers. In addition, the present invention selects a convolution step size s1 that is larger than the convolution step size 1 used in the field of image recognition to reduce the length of the extracted features. A maximum pooling layer with a pool length of 2 is used to capture more concise and locally invariant features. Finally, the denoising layer returns a feature F with a shape of (c1×l1) d , where l1=(l-k1) / 2s1+1.

[0076] MSC Block

[0077] The MSC block contains two processes: the multi-scale feature learning process and the multi-scale feature weighting process, such as Figure 1 shown.

[0078] In the MSC block, the feature Fd extracted by the denoising layer is fed back to n parallel Conv1D layers. The kernels of the n Conv1D layers have different widths, denoted as k2i, (i=1,2,…,n). Kernels with different widths act as filters with different frequency domain resolution scales to simultaneously extract features of different frequency bands of the input signal.

[0079] It is worth noting that each of the n Conv1D layers has .

[0080] All Conv1D layers have the same convolution stride, which is set to 1.

[0081] Each Conv1D layer in the MSC block can extract a feature map at one scale, and the extracted feature map is represented as F Si In order to make the features of each scale have the same length, a padding strategy is adopted here. Si The shape of is (c2'×l2), where l2=l1. In addition, considering that the features of different channels at different scales have different fault sensitivity levels, and the fault sensitivity level of each channel may also change under different operating conditions and different health conditions, an adaptive weight vector W is introduced into the MSC block to dynamically modulate the multi-scale features. By combining features at multiple scales, a multi-scale feature map Fm is formed, which is expressed as:

[0082] F m =[F S1 ; F S2 ;…;F Sn ] (15)

[0083] F m The shape of is (c2×l2), where c2=nc2'. The connection layer acts as a collector that can aggregate features of all scales to form a multi-scale feature set, so the connection layer can retain all sensitive features from kernel convolution layers of different scales. After that, a time-distributed Conv1D layer (TDConv1D) is applied to F m , and returns a feature V with a shape of (c2×1). The structure of TDConv1D is as follows Figure 2 shown.

[0084] In TDConv1D, the Conv1D layer is applied to both F m Each feature channel of tdconv1d is used to learn the global information of each channel. The channel and kernel width of the TDConv1D layer are set to 1 and l2 respectively. Finally, F m The c-th channel in is compressed into the c-th element in the feature V. The weight vector W is calculated by the following equation:

[0085] W=σ(W2f(W1V+b1)+b2) (16)

[0086] Where σ refers to the sigmoid activation, which can implement the gating mechanism. The weight vector W is a 1D vector with a shape of (c2×1). The cth element in W represents F m The fault sensitivity level of the c-th channel characteristic.

[0087] The weight vector W is composed of the model parameters and F m After the MSCNN model is well trained, the model parameters are fixed, so the weight vector W only changes with F m Change, that is, the weight vector W changes with the model input x. Therefore, the weight vector W has better adaptability than a fixed vector. Finally, the multi-scale feature Fm is modulated in the following way:

[0088] F′ m =F m ·W (17)

[0089] Where “·” represents channel multiplication operation.

[0090] Since W can be dynamically adjusted by the model according to the input signal, the multi-scale feature F can be adaptively modulated m To obtain the optimized multi-scale feature F m ', F m ' and F m have the same shape. Each element in W is a value between 0 and 1. The weight vector W acts as a gate that emphasizes fault-sensitive features and suppresses features that are sensitive to operating conditions. Therefore, F m 'More sensitive to faults and less sensitive to changes in operating conditions.

[0091] Multi-scale feature fusion layer

[0092] After the MSC block, a multi-scale feature fusion layer is used to extract the optimal multi-scale features F m 'The multi-scale fusion features are further extracted in . The goal of this layer is to learn the dependencies between features of different scales and channels. The multi-scale feature fusion layer is actually a Conv1D layer with parameters c3@k3 and stride s3. A maximum pooling layer with a pooling length of 2 is used to reduce the dimension of the features. This layer returns a feature F with a shape of (c3×l3) f , where l3 = (l2-k3) / 2s3+1.

[0093] Classification layer

[0094] Finally, feature F f is flattened into a 1-dimensional feature vector and fed to the FC layer and the softmax layer. The softmax layer is defined as:

[0095]

[0096] Where m is the label and M is the total number of labels. θ represents the parameters of the softmax layer.

[0097] Cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network

[0098] Aiming at the problem of low diagnostic accuracy of traditional methods in cable insulation diagnosis and identification, a cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network (ACEEMDAN-MSCNN) is proposed by combining the advantages of CEEMDAN and MSCNN. The implementation process of the model is as follows:

[0099] (1) The sliding window technology is used to decompose the original fault signal into several equal sub-signals, each of which is used as an independent sample. Subsequently, the training set and the validation set are selected according to the random principle to ensure that the data samples of the fault type account for half of each in the training set and the validation set, and the proportion is balanced.

[0100] (2) Reconstruct the training set fault data and the validation set fault data based on adaptive CEEMDAN;

[0101] (3) The reconstructed training set fault data is input into MSCNN for training to obtain relevant parameters with optimal training results.

[0102] (4) Input the reconstructed verification set fault data into the trained MSCNN to obtain the fault recognition result.

[0103] Experimental results and comparative analysis

[0104] Test data

[0105] Fault simulation tests were performed to collect and analyze current readings to verify the fault detection algorithm proposed in the present invention. The current waveform analysis tool used was Matlab 2020a version; the technical basis for building the model was the Pytorch framework.

[0106] The current waveform of a high-voltage cable in a coal mine was selected as the experimental sample, covering the current waveform of the cable under four different working conditions, namely normal working condition, series fault working condition, parallel fault working condition and mixed fault working condition, with 4000 data samples under each working condition. The collected data was standardized and the fault sample data was divided into two groups.

[0107] Model parameters

[0108] The present invention selects model parameters through cross-validation experiments. In the ACEEMDAN-MSCNN model, a 4@64 convolutional 1D layer with a stride of 8 is used in the denoising layer of MSCNN, the multi-scale feature learning process in the AWMSC block consists of 4@2, 4@4, 4@6 and 4@8 Conv1D layers with a stride of 1, the multi-scale feature weighting process consists of a 16 (1@636) TDConv1D layer and two FC layers with 6 and 16 neurons, and a 4@8 convolutional 1D layer with a stride of 4 is used in the multi-scale feature fusion layer. The batch size is 50, the number of epochs is 50, and the optimization algorithm is Adam.

[0109] Comparison between ACEEMDAN-MSCNN and CEEMDAN-MSCNN, EMD-MSCNN, and MSCNN:

[0110] The ACEEMDAN-MSCNN is compared with CEEMDAN-MSCNN, EMD-MSCNN, and MSCNN. The loss function curves of the four models and the accuracy curves of the recognition results are shown in Figure 2. Figure 3 As shown in the figure, it can be seen that the diagnostic accuracy of the VMD-CMDE-MCDBN model is significantly higher than that of the other three models, and the diagnostic accuracy converges to the maximum value, which greatly improves both performance and accuracy.

[0111] Figure 4 ,Table 1 shows the fault recognition results of the four models.,It can be seen that the recognition accuracy of the VMD-CMDE-MCDBN model has been improved to a certain extent, and the,performance is more stable, which fully demonstrates the effectiveness of combining VMD with CMDE.

[0112] Table 1 Fault identification results of different models

[0113]

[0114] Noise robustness comparison

[0115] In order to simulate the noise environment, the present invention injects additive Gaussian white noise into the original fault signal to construct noise signals with different signal-to-noise ratios (SNR). The signal-to-noise ratio is defined as:

[0116]

[0117] Where P signal and P noise denote the power of the original signal and the injected noise respectively.

[0118] The comparison results of the noise robustness of different models are shown in Figure 5As shown. Obviously, the proposed ACEEDAN-MSCNN outperforms other models, with an accuracy of more than 70% at each signal-to-noise ratio. When the signal-to-noise ratio exceeds 0dB, ACEEDAN-MSCNN can achieve an accuracy of more than 94%. By comparing the performance of AWMSCNN and AWMSCNN-II, it can be concluded that the adaptive multi-scale feature weighting process in ACEEDAN-MSCNN can improve the noise resistance of the model. In addition, all multi-scale models outperform CNN, which shows that multi-scale feature learning can extract more robust features from noisy input signals than single-scale CNN. The above results show that the proposed ACEEDAN-MSCNN can be used for fault diagnosis under variable speed in actual noisy industrial environments.

[0119] Comparison with existing models

[0120] In order to verify the recognition performance of the VMD-CMDE-MCDBN model and the current mainstream fault diagnosis algorithms, 1DCNN-LSTM, CNN-BES-ELM, CNN-A-BiLSTM, IGOA-KELM, and AFF-Stablenet are used as comparison algorithms for comparative experiments. The experimental results are shown in Table 2.

[0121] Table 2 Comparison of fault diagnosis results of different models

[0122]

[0123]

[0124] As can be seen from Table 2, the ACEEDAN-MSCNN model has better diagnostic performance than other models. This is mainly because: (1) The IGOA-KELM and AFF-Stablenet models use machine learning methods to perform the final fault identification, and can well identify the deep features of the signal data; (2) The 1DCNN-LSTM, EWT-CNN-BES-ELM, and CNN-A-BiLSTM models use deep learning methods for feature extraction and model training. Although they have achieved certain results, due to the complexity of the fault signal characteristics, they cannot learn rich and complementary feature information at different scales, making their fault diagnosis results significantly lower than those of the ACEEDAN-MSCNN model.

[0125] in conclusion

[0126] This paper focuses on the fault diagnosis problem of XLPE cables and proposes an XLPE cable insulation fault diagnosis model (ACEEMDAN-MSCNN) based on adaptive CEEMDAN and multi-scale convolutional neural network. The model first uses CEEMDAN decomposition to perform multi-scale analysis on each data sequence, then adaptively selects effective IMF components as feature factors, and finally calculates the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension; then, a multi-scale convolutional neural network model, MSCNN, is proposed, which takes the reconstructed fault signal as input, without any manual features, and can automatically extract fault-sensitive features and perform fault classification. In the experiment, ACEEMDAN-MSCNN outperforms the basic single-scale CNN model and other multi-scale models.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network, characterized in that: include: The sliding window technology is used to decompose the original fault signal into several equal sub-signals, each of which is regarded as an independent sample; Then, the training set and validation set are selected according to the random principle to ensure that the data samples of the fault type account for half of the training set and the validation set respectively, and the proportion is balanced; Reconstruct the training set fault data and the validation set fault data based on adaptive CEEMDAN; The reconstructed training set fault data is input into MSCNN for training to obtain relevant parameters with optimal training results; The reconstructed verification set fault data is input into the trained multi-scale convolutional neural network MSCNN to obtain the fault recognition result.

2. The XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network according to claim 1, characterized in that: The reconstruction of the training set fault data and the verification set fault data based on the adaptive CEEMDAN includes: Use CEEMDAN decomposition to perform multi-scale analysis on each data sequence, then adaptively select effective IMF components as characteristic factors, and finally calculate the corresponding energy entropy as the final eigenvalue to reconstruct all data samples with the same dimension. represents the jth feature variable in the i-th data sample Yi, and f is the dimension of Yi.

3. The XLPE cable insulation fault diagnosis model based on adaptive CEEMDAN and multi-scale convolutional neural network according to claim 1, characterized in that: The input of the multi-scale convolutional neural network MSCNN is the reconstructed fault data, and the output is the predicted label indicating the fault condition; the input sample X∈Rl is represented as X={x1,x2,…,xl}, where l is the length of the input sample, and the output Y is a One-Hot vector. The fault diagnosis task is defined as using the MSCNN model to obtain the label Y based on the reconstructed fault diagnosis data X. The MSCNN model has four parts: a denoising layer, an AWMSC block, a multi-scale feature fusion layer, and a classification layer.

Images