Cable insulation fault identification model based on VMD-CMDE and multi-scale cascade deep belief network

Abstract

The invention discloses a cable insulation fault identification model (VMD-CMDE-MCDBN) based on a VMD-CMDE and a multi-scale cascade deep belief network. The method comprises the following steps: firstly, decomposing original cable fault sample data by using variational mode decomposition (VMD), and calculating the kurtosis of a mode component obtained by decomposition; then fault signal reconstruction is carried out according to the kurtosis, a composite multi-scale walk entropy (CMDE) of a reconstructed signal is calculated, and feature data of fault sample data is obtained; and finally, taking the fault sample feature data as the input of a multi-scale cascaded deep belief network (MCDBN) to obtain a final recognition result. According to the model, the advantages of VMD and CMDE are fully utilized, rich and complementary feature information is learned on different scales, and better classification performance is achieved compared with an existing model.

Description

Technical Field

[0001] The present invention relates to cable fault technology, and in particular to a cable insulation fault identification model based on VMD-CMDE and a multi-scale cascaded deep belief network. Background Art

[0002] At present, with the continuous improvement of China's power supply system, power cables have been widely used in various fields. However, cables work in a humid underground environment for a long time, which is prone to short-circuit faults, which may cause damage to power equipment and paralysis of the power supply network. Therefore, it is very important to fully understand the characteristics of cable faults, study cable fault diagnosis methods, and accurately and quickly diagnose the type of cable faults for the stable and reliable operation of the power grid. In the initial projects, offline detection methods are often used to determine the type of cable faults, but this method is cumbersome and time-consuming, requires power outages for maintenance, and will also affect the power consumption in non-fault areas.

[0003] With the development of power cable fault diagnosis technology, neural network technology has gradually been applied to cable fault diagnosis. Hu Yufeng et al. built a cable simulation model through Simulink, and then extracted the voltage signal characteristics of the normal and faulty states of the cable respectively. Finally, they used artificial neural networks to diagnose cable faults and were able to identify cable fault state signals. In recent years, experts and scholars have proposed a method to judge the fault line of the circuit by zero-sequence current. For example, Xue Fei et al. decomposed the zero-sequence current of each line through wavelet packet transform, converted the fault line selection into a multi-classification problem, and successfully predicted the fault line. This method extracts the characteristic quantity of zero-sequence current as the basis for judging cable faults, but for actual power grid lines, there will be factors such as the current compensation effect of arc suppression coils and the influence of grounding resistance, resulting in low classification accuracy. In order to solve the problem of low accuracy of underground cable diagnosis, Feng Yin et al. proposed a cable fault diagnosis method based on EMD, Hilbert transform and SVM. By extracting the time-frequency attributes of the signal, the support vector machine algorithm was used to achieve the classification of cable faults. Some people use LMD multi-scale approximate entropy to extract fault features. Some people use the variational mode decomposition (VMD) method to calculate the gear signal, and after decomposition, select four modal components to calculate the permutation entropy and extract features. Some people significantly improve fault identification by combining the empirical wavelet transform (EWT) with various dispersion entropy (DE) algorithms.

[0004] In recent years, deep learning theory has promoted the development of artificial intelligence. Its effective training method and special network structure can better mine effective features in data that are conducive to identification. Compared with traditional machine learning, deep learning has higher accuracy in image processing, signal recognition, etc., and has better generalization performance. After studying the characteristics of early cable faults, some people used convolutional neural networks to identify various cable faults. Some people studied the detection method of cables in non-operating states, used deep learning algorithms to reduce noise on the sound, calculated the logarithmic power spectrum of cable fault signals, and used it as data for network training, solving the shortcomings of machine learning in sound noise reduction processing. Some people found that training mechanical vibration signals of related faults through deep learning neural networks is more conducive to fault identification and classification, and pointed out the advantages of using deep learning theory for fault diagnosis, which is mainly reflected in breaking the researchers' dependence on multiple signal processing technologies and fault diagnosis experience. Starting from the statistical characteristics of vibration signals, some people achieved simultaneous recognition of different types and degrees, and finally obtained higher classification accuracy. The results show that the application of DBN in fault diagnosis has better results than traditional fault diagnosis.

[0005] Although the above work has achieved good results, the following problems still exist: Currently, in terms of cable fault diagnosis, there is relatively little research on the intelligent diagnosis of cable short-circuit faults; the boundary between the electrical characteristics of fault conditions and normal conditions is relatively vague, resulting in low recognition accuracy and difficulty in meeting actual needs; the current signal is a time series data, and the output at the current time is related to the previous and subsequent states, and the current algorithm does not take this into account.

[0006] In view of the above problems, the present invention proposes a cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network (VMD-CMDE-MCDBN). Summary of the invention

[0007] The main purpose of the present invention is to provide a cable insulation fault identification model based on VMD-CMDE and a multi-scale cascaded deep belief network, integrate the improved multi-scale coarse-grained process into the DBN architecture, inject the multi-scale self-learning capability into the traditional DBN in a parallel manner, construct a multi-scale cascaded deep belief network MCDNB, and use the fault signal feature data processed by VMD and CMDE as the input of MCDNB for cable insulation fault identification.

[0008] The technical solution adopted by the present invention is: a cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network, including:

[0009] The original vibration signal is divided into a series of sub-signals of equal size through a sliding window, and each sub-signal is regarded as a sample. After obtaining the sub-signals, the training samples and test samples are selected according to the random standard, and the data samples of each fault mode are randomly assigned to the training / test samples in a ratio of 1:1;

[0010] Perform VMD decomposition and preprocessing on the training sample and test sample data, calculate the kurtosis of the decomposed modal components and sort them;

[0011] Select the first three modal components to reconstruct the training sample and test sample data, calculate the composite multi-scale scatter entropy of the reconstructed training sample and test sample data, and obtain the eigenvalues ​​of the training sample and test sample data;

[0012] The characteristic values ​​of the training samples are input into the MCDNB model for training, and the relevant parameters with the optimal training results are obtained. The scale is set to 3 and the number of RBMs is set to 3;

[0013] Input the test sample data into the trained MCDNB to obtain the fault recognition result.

[0014] Furthermore, the multi-scale cascaded deep belief network includes an improved multi-scale coarse-grained process and an MCDNB model.

[0015] Furthermore, the improved multi-scale coarse-grained process includes:

[0016] For a given time series {x(i):1≤i≤N}, the improved multi-scale coarse-grained time series with a scaling factor of τ is obtained by the following formula (19):

[0017]

[0018] Where τ = 1, 2, ... is the scaling factor, and N is the length of the original time series. According to formula (19), the original signal is divided into τ coarse-grained time series containing different and complementary fault features, and the length of the coarse-grained time series changes very slowly with the increase of the scaling factor. For the τth coarse-grained time series, the length is equal to N-τ+1.

[0019] Furthermore, the MCDNB model includes:

[0020] Improved multi-scale coarse-grained layer:

[0021] This layer performs improved multi-scale coarse-grained processing on the original signal to obtain coarse-grained time series at different scales. When the scale factor τ = 1, 2, and 3, three improved coarse-grained time series are obtained from the original signal ( and ) to capture multi-scale information in multiple scale ranges;

[0022] Multi-scale feature learning layer:

[0023] This layer automatically learns a high-level and effective feature representation from multiple improved coarse-grained time series in a parallel manner through multiple pairs of stacked RBM learning layers. Figure 3 It can be clearly seen that the three improved coarse-grained time series ( and ) is input into a DBN with three RBMs in parallel. Assuming {x(i):1≤i≤N} is an input original signal of length N, the three feature learning vectors of the RBM3 hidden layer can be obtained in parallel through three stacked RBM3s. Among them, the output feature of the jth node in the RBM3 hidden layer is It is expressed as:

[0024]

[0025] where σ(·) represents the sigmoid activation function, represents the link weight between the ith visible element and the jth hidden element, represents the deviation of the jth hidden element, The jth element of the improved coarse-grained time series has a length of N-τ+1. Taking the scale factor τ=3 as an example, the improved coarse-grained time series The output of the j node in the hidden layer of RBM3 is defined as follows:

[0026]

[0027] In a cascaded manner, the representation of high-level features is as follows:

[0028]

[0029] Compared with the traditional single-scale feature representation, the above high-level feature representation has more comprehensive and richer fault features, which are learned from multiple scales of the original signal, so that MCDNB can enhance the feature learning ability and improve the discrimination between categories.

[0030] Classification layer:

[0031] The softmax classifier is used instead of BPNN to output the probability distribution of each fault mode. Assuming that the input sample x contains k types of health states, the probability corresponding to the jth class can be calculated by formula (23):

[0032]

[0033] Where θ is the parameter learned from DBN.

[0034] Advantages of the present invention:

[0035] The cable insulation fault identification model based on VMD-CMDE and multi-scale cascade deep belief network (VMD-CMDE-MCDBN) of the present invention. First, the original cable fault sample data is decomposed by variational mode decomposition (VMD), and the kurtosis of the modal component obtained by decomposition is calculated; then the fault signal is reconstructed according to the kurtosis, and the composite multi-scale distribution entropy (CMDE) of the reconstructed signal is calculated to obtain the feature data of the fault sample data; finally, the fault sample feature data is used as the input of the multi-scale cascade deep belief network (MCDBN) to obtain the final recognition result. Experimental results show that the VMD-CMDE-MCDBN model makes full use of the advantages of VMD and CMDE, learns rich and complementary feature information at different scales, and achieves better classification performance than existing models.

[0036] 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

[0037] 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.

[0038] Figure 1 It is a multi-scale coarse-grained process when the scale factor τ = 2 and τ = 3;

[0039] Figure 2 It is an improved multi-scale coarse-grained process when the scale factor τ = 2 and τ = 3;

[0040] Figure 3 It is the overall framework of the MCDNB model;

[0041] Figure 4 It is data segmentation;

[0042] Figure 5 This is the center frequency curve when 5K=4;

[0043] Figure 6 It is the center frequency curve of K=5;

[0044] Figure 7 This is the center frequency curve when K=6;

[0045] Figure 8 It is the change curve of composite multi-scale dispersion entropy under different scale factors;

[0046] Fig. 9 is the fault diagnosis error variation curve of different models;

[0047] Fig.10 is the fault diagnosis accuracy variation curve of different models;

[0048] Fig.11 This is a comparison chart of the Accuracy values ​​of the four models;

[0049] Fig.12 This is a comparison chart of the Precision values ​​of the four models;

[0050] Fig.13 This is a comparison chart of the recall values ​​of the four models;

[0051] Fig.14 This is a comparison chart of the F1-Score values ​​of the four models. DETAILED DESCRIPTION

[0052] 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.

[0053] refer to Figures 1 to 14 , a cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network is provided.

[0054] Variational Mode Decomposition (VMD) algorithm:

[0055] The essence of the variational mode decomposition (VMD) algorithm is to select the number of components (parameter K) and decompose the original signal f(t) into the corresponding number of sub-signal components u k ; These decomposed modal components can ensure the sparsity and reproducibility of the input signal. In summary, the bandwidth is estimated using Gaussian smoothing of the demodulated signal, and then the constraints are divided into the following parts:

[0056]

[0057] The optimal solution of most constraint models is solved by the alternating direction method of multipliers (ADMM), which updates To find the Lagrangian augmented "saddle point"; the specific steps are as follows:

[0058] First initialize λ'; n←O; when k=1, let n←n+1, so that k is updated to u k

[0059]

[0060] Update k :

[0061]

[0062] Repeat equations (3) and (4) until the following iteration conditions are met:

[0063]

[0064] Generally speaking, The problem is transformed into a minimum problem. Similarly, the center frequency We can get:

[0065]

[0066] Composite Multiscale Scattering Entropy (CMDE):

[0067] Divided Entropy (DE):

[0068] The diffusion entropy (DE) is an indicator to measure the complexity of time series. The main steps of constructing DE are as follows:

[0069] Assume that the time series x of length N is i , i=1,2,...,N}Use the normal distribution function (7) to map the time series x to y={y j ,j=1,2,...,N},y j ∈(0,1)

[0070]

[0071] where μ is the mathematical expectation, σ 2 is the variance.

[0072] Use equation (8) to perform linear transformation and map y to [1,2,…c]:

[0073]

[0074] Where R is the integral function and c is the number of categories.

[0075] Calculate the embedding vector As shown below:

[0076]

[0077] where m is the embedding dimension and d is the time delay.

[0078] Calculate the distribution mode Probability for:

[0079]

[0080] in express arrive The number of mappings.

[0081] The DE value of the original signal x is:

[0082]

[0083] Composite Multiscale Scattering Entropy (CMDE):

[0084] The composite multi-scale scatter entropy is obtained by multi-scale optimization based on the scatter entropy. For the initial time series {u(i), i=1,2,...,L}, when the time series is in the kth coarsening and the scale factor is τ, It can be given by formula (12):

[0085]

[0086] The CMDE of τ under the scaling factor is defined as:

[0087]

[0088] Deep Belief Network (DBN):

[0089] As one of the typical deep learning algorithms, the Deep Belief Network (DBN) has a good development prospect in the field of fault identification. The Deep Belief Network (DBN) is a probabilistic artificial neural network with multiple hidden layers, which is composed of multiple Restricted Boltzmann Machines (RBM) stacked together. Through the Restricted Boltzmann Machine, the relevant function can be obtained as follows:

[0090]

[0091] Where θ is the node parameter of the restricted Boltzmann machine, θ = {W ij ,a i ,b j} are all real numbers, a i represents the offset coefficient of visible element i; Wij represents the weight value of hidden element j and visible element i; b j represents the offset coefficient of latent element j.

[0092] When these parameters are constant, the joint probability distribution can be obtained according to the function, as shown in formula (15):

[0093]

[0094] Where Z(θ) is the normalization factor; a i and b i is the offset coefficient; h j 、v i are the state variables of hidden and explicit elements; W ijare the weights of hidden and visible units.

[0095] From the special structure of the energy function, we can see that the layers of RBM are connected. When the state of the hidden layer is known, the conditions for the activation of different visible elements are independent. The probability of visible element activation is shown in formula (16):

[0096] P(vi=1|h,θ)=σ(ai+∑jWijhj) (16)

[0097] Similarly, the activation probability of the hidden unit is:

[0098] P(h i =1|v,θ)=σ(b j +∑ i v i W ij ) (17)

[0099] in is the Sigmoid activation function.

[0100] Multi-scale Cascaded Deep Belief Networks:

[0101] This application proposes a multi-scale cascaded deep belief network, the MCDBN model, which integrates the improved multi-scale coarse-grained process into the traditional DBN, so as to learn rich and complementary feature information at different scales to achieve better classification performance.

[0102] Improved multi-scale coarse-grained process:

[0103] At present, a multi-scale coarse-grained process is introduced into the traditional multi-scale analysis method to capture the fault characteristics of the original signal of cable insulation at different time scales. Figure 1 The traditional coarse-grained process with scaling factors τ = 2 and τ = 3 is shown. For a given time series {x(i): 1≤i≤N}, the factor of the coarse-grained time series τ at a certain scale is calculated by formula (18):

[0104]

[0105] Where τ = 1, 2, ... is a scaling factor. When τ = 1, the coarse-grained time series obtained is essentially the original signal of cable insulation. When τ > 1, the signal is divided into τ coarse-grained time series with a length of

[0106] from Figure 1It can be clearly seen from equation (18) that the traditional coarse-grained method realizes the multi-scale operation of the original signal through downsampling and smoothing of non-overlapping windows, but lacks continuous shift operation, which means that some inherent feature information in the original signal may be ignored. In addition, for the traditional coarse-grained process, the length of the coarse-grained time series decreases exponentially with the increase of the scale factor. In order to solve this problem, the present invention proposes an improved multi-scale coarse-grained process to obtain the feature representation of the original signal at multiple time scales. Figure 2 The improved multi-scale coarse-grained process with scaling factor τ = 2 and τ = 3 is shown. For a given time series {x(i): 1≤i≤N}, the improved multi-scale coarse-grained time series with scaling factor τ can be obtained by the following equation (19).

[0107]

[0108] Where τ = 1, 2, ... is the scaling factor, and N is the length of the original time series. According to formula (19), the original signal can be divided into τ coarse-grained time series containing different and complementary fault features, and the length of the coarse-grained time series changes very slowly with the increase of the scale factor. For the τth coarse-grained time series, the length is equal to N-τ+1. Therefore, when the improved multi-scale coarse-grained process is introduced to characterize the original signal within a certain scale range, a more accurate and reliable feature information estimation can be achieved. In fact, the improved multi-scale coarse-grained process is essentially equivalent to a compact moving average process of overlapping adjacent data. Compared with some existing multi-scale transformation methods, the improved multi-scale coarse-grained process has the advantages of being easier to implement and having less running time, thereby simplifying the model complexity and improving the efficiency of feature learning.

[0109] MCDBN model:

[0110] The MCDBN model architecture mainly consists of three layers: (1) improved multi-scale coarse-grained layer; (2) multi-scale feature learning layer; (3) classification layer, such as Figure 3 shown.

[0111] Improved multi-scale coarse-grained layer:

[0112] This layer performs improved multi-scale coarse-grained processing on the original signal to obtain coarse-grained time series at different scales. The specific calculation process is shown in formula (19). Figure 3 As shown in Figure 2, when the scale factor τ = 1, 2, and 3, three improved coarse-grained time series are obtained from the original signal ( and ) to capture multi-scale information in multiple scale ranges.

[0113] Multi-scale feature learning layer:

[0114] This layer automatically learns a high-level and effective feature representation from multiple improved coarse-grained time series in a parallel manner through multiple pairs of stacked RBM learning layers. Figure 3 It can be clearly seen that the three improved coarse-grained time series ( and ) is input into a DBN with three RBMs in parallel. Assuming {x(i):1≤i≤N} is an input original signal of length N, the three feature learning vectors of the RBM3 hidden layer can be obtained in parallel through three stacked RBM3s. Among them, the output feature of the jth node in the RBM3 hidden layer is It is expressed as:

[0115]

[0116] where σ(·) represents the sigmoid activation function, represents the link weight between the ith visible element and the jth hidden element, represents the deviation of the jth hidden element, The jth element of the improved coarse-grained time series has a length of N-τ+1. Taking the scale factor τ=3 as an example, the improved coarse-grained time series The output of the j node in the hidden layer of RBM3 is defined as follows:

[0117]

[0118] In a cascaded manner, the representation of high-level features is as follows:

[0119]

[0120] Compared with the traditional single-scale feature representation, the above high-level feature representation has more comprehensive and richer fault features, which are learned from multiple scales of the original signal, so that MCDNB can enhance the feature learning ability and improve the discrimination between categories.

[0121] Classification layer:

[0122] In traditional DBN, BP neural network is regarded as the classification tool of the last layer. Although it has the advantages of self-learning and nonlinear mapping, it also has the problems of local minimization and slow convergence. Therefore, the present invention adopts softmax classifier instead of BPNN to output the probability distribution of each fault mode. Assuming that the input sample x contains k types of health states, the probability corresponding to the jth class can be calculated by formula (23):

[0123]

[0124] Where θ is the parameter learned from DBN.

[0125] Cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network:

[0126] Aiming at the problem that the traditional cable insulation fault identification model has cumbersome processing and low accuracy, this paper combines the advantages of VMD, CMDE and MCDNB, and proposes a cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network (VMD-CMDE-MCDBN). The implementation process of the model is as follows:

[0127] (1) The original vibration signal is divided into a series of sub-signals of equal size through a sliding window, and each sub-signal is regarded as a sample. The data segmentation diagram is shown in the figure. Figure 4 As shown. After obtaining the sub-signals, the training samples and test samples are selected according to the random criteria. Since the training / test data percentage has a great impact on the diagnosis results, it needs to be selected carefully. The more training data, the better the classification effect, but the more training time will be required. Therefore, in order to strike a balance between accuracy and efficiency, the data samples of each fault mode are randomly assigned to the training / test samples in a 1:1 ratio.

[0128] (2) Perform VMD decomposition and preprocessing on the training sample and test sample data, calculate the kurtosis of the decomposed modal components and sort them;

[0129] (3) Select the first three modal components to reconstruct the training sample and test sample data, calculate the composite multi-scale scatter entropy of the reconstructed training sample and test sample data, and obtain the eigenvalues ​​of the training sample and test sample data;

[0130] (4) Input the characteristic values ​​of the training samples into the MCDNB model for training to obtain relevant parameters with optimal training results. When the MCDNB model executes the multi-scale coarse-grained process, appropriate scale parameters should be preset. Based on the prior knowledge of multi-scale analysis, it is known that the more scales / RBMs, the longer the training time and the lower the computational efficiency. Therefore, without loss of generality, the present invention sets the scale to 3 and the number of RBMs to 3.

[0131] (5) Input the test sample data into the trained MCDNB to obtain the fault recognition result.

[0132] Experimental results and comparative analysis:

[0133] Experimental data:

[0134] The present invention takes a single-core 35kV cross-linked polyethylene (XLPE) power cable with a cross-sectional area of ​​46mm2 as the research object, and collects the fault current signal on the cable line, and the fault signal sampling time is 0.2s. During sampling, four types of cable fault signals were collected, each type of fault contains 10,000, and the total number of sample samples is 40,000. The number of samples collected for each type of signal is shown in Table 1. Wherein I indicates that a ground fault occurs in phase A, II indicates that a ground fault occurs in phases A and B, III indicates that a ground fault occurs in phases A, B, and C, and IV indicates that a short circuit fault occurs between phases A and B.

[0135] Table 1 Fault sample sampling quantity

[0136]

[0137] Before the experiment, the collected fault data is first normalized using formula (24).

[0138]

[0139] Model parameter selection:

[0140] The parameters of the VMD-CMDE-MCDBN model are set as follows: the number of iterations is set to 500; the length parameter N in CMDE is 1024, the dimension M is 3, the number of categories C is 6, and the delay D is 1; the initial learning rate of DBN is set to 0.01, the gradient decay factor is set to 0.8, and the convergence factor is set to 25. Since the overall performance of the model is crucial, the present invention analyzes the values ​​of the modal component K in VMD and the scale factor τ in CMDE respectively.

[0141] Modal component K:

[0142] The present invention takes the center frequency of the modal component of the fault signal as an example. The modal component K is reflected in the number of curves in the center frequency diagram. The K value is determined by observing the convergence trend of the curve. When K = 4, K = 5, and K = 6, the corresponding center frequency curves are as follows: Figure 5 , 6 , 7. The horizontal axis in the figure represents the number of iterations, the vertical axis represents the center frequency, and the four curves represent the center frequency convergence process of the four modal components respectively.

[0143] As can be seen from the figure, when K = 4, the four curves do not overlap, proving that there is no modal aliasing. Although there are occasional fluctuations in the previous iteration, the convergence speed is fast. When K is 5, the center frequencies corresponding to the five modal components converge smoothly, with small fluctuations, and there is no intersection of the curves. When K = 6, it can be clearly seen from the curves corresponding to the six modal components that the third, fourth, and fifth-order curves also correspond to the intersection of the third, fourth, and fifth-order modal components, respectively, which proves that there is modal mixing between the modal components and the convergence speed is slow. In summary, the preset value of the modal component is 5, which is more effective for the decomposition effect of the signal and is conducive to the next step of feature extraction.

[0144] Scaling factor τ:

[0145] The present invention determines the optimal value of the composite multi-scale dispersion entropy by comparing the changes of the composite multi-scale dispersion entropy under different scale factors τ. Figure 8 The random point entropy curves corresponding to scale factors from 1 to 12 are shown. The horizontal axis is the number of scale factors, and the vertical axis is the composite multi-scale scatter entropy.

[0146] from Figure 8 It can be seen that, except for the type I fault signal, the overall trend of the signals of the other three faults is to first decrease and then flatten out; in the process of the scale factor changing from 1 to 4, except for the upward trend under normal circumstances, the other three fault signals all show a downward trend, and the downward trend is obvious from the instantaneous rate of change; when the scale factor is 4 to 8, the overall decrease is relatively gentle, with occasional fluctuations, and the downward trend of type II fault is more obvious; when the scale factor is in the range of 8 to 10, the decrease is gentle, and the entropy of the fault signal slowly approaches. The reason why the type I fault signal is different from the other three fault signals is that there is no periodic vibration similar to the fault signal; when the scale factor is 10 to 12, the entropy of the fault signal tends to overlap, and the CMDE value does not change much, but as the parameters increase, the simulation time becomes longer. In summary, when the scale factor is 10, it can not only ensure that the deep information of the vibration signal is extracted, but also ensure that the time is not excessively consumed. Therefore, the scale factor of the composite multi-scale spread entropy of the present invention is set to 10.

[0147] Evaluation indicators:

[0148] The evaluation indicators used for cable insulation fault identification of the present invention are: Accuracy, Precision, Recall, F1-Score, and Table 4 is a confusion matrix.

[0149] Table 4 Confusion matrix

[0150]

[0151]

[0152] Comparison between VMD-CMDE-MCDBN, VMD-MCDBN, CMDE-MCDBN and MCDNB:

[0153] In order to verify the effectiveness of the combination of VMD and CMDE in the VMD-CMDE-MCDBN model, the present invention compares VMD-CMDE-MCDBN with VMD-MCDBN, CMDE-MCDBN, and MCDNB. The loss function curves of the four models and the accuracy curves of the recognition results are shown in Figure 2. Fig. 9 , Fig.10 As shown in the figure, it can be seen that the diagnosis accuracy of the VMD-CMDE-MCDBN model is significantly higher than that of the other three models, and the diagnosis error converges to a smaller value, which greatly improves both performance and accuracy.

[0154] Table 2 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.

[0155] Table 2 Fault identification effects of different models

[0156]

[0157] Comparison between VMD-CMDE-MCDB and EMD-CMDE-MCDBN, EEMD-CMDE-MCDBN and VMD-DE-MCDBN:

[0158] In order to verify the effectiveness of single VMD and single CMDE in the VMD-CMDE-MCDBN model, the present invention compares VMD-CMDE-MCDBN with EMD-CMDE-MCDBN, EEMD-CMDE-MCDBN, and VMD-DE-MCDBN. The diagnostic results of the four models are shown in Table 3. It can be seen that the VMD-CMDE-MCDBN model has a certain degree of improvement in the four evaluation indicators compared with the other three models, indicating that the use of VMD for signal decomposition in the VMD-CMDE-MCDBN model is more effective than the use of EMD and EEMD for signal decomposition, and the use of CMDE for feature extraction can extract better features than the use of DE model, effectively improving the fault identification accuracy of the model.

[0159] Table 3 Fault identification effects of different models

[0160]

[0161] Compared with existing models:

[0162] In order to verify the recognition performance of the VMD-CMDE-MCDBN model and the current mainstream fault diagnosis algorithms, POA-VMD-RF, EWT-SSA-SVM, CWT-CNN, EEMD-IGWO-SVM, and CNN-SVM are used as comparison algorithms for comparative experiments. The experimental results are shown in Table 5. Figure 11-Figure 14 shown.

[0163] Table 5 Comparison of fault diagnosis results of 6 models

[0164]

[0165] It can be seen from the above charts that the VMD-CMDE-MCDBN model has better performance in all four evaluation indicators than other models. This is mainly because: (1) The three models of POA-VMD-RF, EWT-SSA-SVM, and EEMD-IGWO-SVM use machine learning methods such as RF and SVM to perform the final fault identification, and can well identify the deep features of the signal data; (2) The CWT-CNN and CNN-SVM 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, the simple CWT or CNN cannot learn rich and complementary feature information at different scales, making their fault diagnosis results significantly lower than the VMD-CMDE-MCDBN model.

[0166] Aiming at the problem of multi-scale feature learning of fault signal data, this paper proposes a cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network (VMD-CMDN-MCDBN). This model integrates the improved multi-scale coarse-grained process into the DBN architecture, injects multi-scale self-learning capability into the traditional DBN in a parallel manner, constructs a multi-scale cascaded deep belief network MCDDN, and uses the fault signal feature data processed by VMD and CMDE as the input of MCDDN for cable insulation fault identification. In the experiment, on the one hand, the VMD-CMDN-MCDBN model was compared with VMD-MCDBN, CMDE-MCDBN, and MCDNB. The results showed that the model gave full play to the advantages of VMD and CMDE and had better fault identification results. On the other hand, the VMD-CMDN-MCDBN model was compared with existing models such as POA-VMD-RF, EWT-SSA-SVM, CWT-CNN, EEMD-IGWO-SVM, and CNN-SVM. The results showed that it had a certain degree of improvement in the four evaluation indicators and was an effective fault identification method.

[0167] The present invention proposes an improved multi-scale coarse-graining process, which performs multi-scale operations on the original signal and can mine inherent feature information across multi-scale time scales; by integrating multi-scale operations into the traditional DBN, a new deep learning network MCDNB is proposed, which can directly learn multi-scale feature representation from the original signal in a parallel manner; and a VMD-CMDE-MCDBN model is proposed, which can greatly enhance the learning ability and improve the recognition accuracy.

[0168] 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. A cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network, characterized by: include: The original vibration signal is divided into a series of sub-signals of equal size through a sliding window, and each sub-signal is regarded as a sample. After obtaining the sub-signals, the training samples and test samples are selected according to the random standard, and the data samples of each fault mode are randomly assigned to the training / test samples in a ratio of 1:1; Perform VMD decomposition and preprocessing on the training sample and test sample data, calculate the kurtosis of the decomposed modal components and sort them; Select the first three modal components to reconstruct the training sample and test sample data, calculate the composite multi-scale scatter entropy of the reconstructed training sample and test sample data, and obtain the eigenvalues ​​of the training sample and test sample data; The characteristic values ​​of the training samples are input into the MCDNB model for training, and the relevant parameters with the optimal training results are obtained. The scale is set to 3 and the number of RBMs is set to 3; The test sample data is input into the trained MCDNB to obtain the fault recognition result.

2. The cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network according to claim 1 is characterized in that: The multi-scale cascaded deep belief network includes an improved multi-scale coarse-grained process and an MCDNB model.

3. The cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network according to claim 2 is characterized in that: The improved multi-scale coarse-grained process includes: For a given time series {x(i):1≤i≤N}, the improved multi-scale coarse-grained time series with a scaling factor of τ is obtained by the following formula (19): Where τ = 1, 2, ... is the scaling factor, and N is the length of the original time series. According to formula (19), the original signal is divided into τ coarse-grained time series containing different and complementary fault features, and the length of the coarse-grained time series changes very slowly with the increase of the scaling factor. For the τth coarse-grained time series, the length is equal to N-τ+1.

4. The cable insulation fault identification model based on VMD-CMDE and multi-scale cascaded deep belief network according to claim 2 is characterized in that: The MCDNB model includes: Improved multi-scale coarse-grained layer: This layer performs improved multi-scale coarse-grained processing on the original signal to obtain coarse-grained time series at different scales. When the scale factor τ = 1, 2, and 3, three improved coarse-grained time series are obtained from the original signal ( and ) to capture multi-scale information in multiple scale ranges; Multi-scale feature learning layer: This layer automatically learns a high-level and effective feature representation from multiple improved coarse-grained time series in parallel through multiple pairs of stacked RBM learning layers; the three improved coarse-grained time series ( and ) is input into a DBN with three RBMs in parallel; assuming that {x(i):1≤i≤N} is an input original signal of length N, the three feature learning vectors of the RBM3 hidden layer can be obtained in parallel through three stacked RBM3 Among them, the output feature of the jth node in the RBM3 hidden layer is It is expressed as: where σ(·) represents the sigmoid activation function, represents the link weight between the ith visible element and the jth hidden element, represents the deviation of the jth hidden element, The jth element of the improved coarse-grained time series has a length of N-τ+1. Taking the scale factor τ=3 as an example, the improved coarse-grained time series The output of the j node in the hidden layer of RBM3 is defined as follows: In a cascaded manner, the representation of high-level features is as follows: Compared with the traditional single-scale feature representation, the above high-level feature representation has more comprehensive and richer fault features, which are learned from multiple scales of the original signal, so that MCDNB can enhance the feature learning ability and improve the discrimination between categories; Classification layer: The softmax classifier is used instead of BPNN to output the probability distribution of each fault mode. Assuming that the input sample x contains k types of health states, the probability corresponding to the jth class can be calculated by formula (23): Where θ is the parameter learned from DBN.

Images