Cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU

Abstract

The invention discloses a cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU. The cable insulation fault diagnosis model comprises the steps that an improved empirical mode decomposition method IEMD is used for decomposing cable fault sample data to obtain an intrinsic mode function IMF; designing a fusion model based on a convolutional neural network CNN and a residual gating cycle unit ResGRU for pre-training; a fuzzy particle swarm optimization algorithm NFPSO based on a neighborhood is designed to optimize and adjust hyper-parameters of a CNN-ResGRU model, and an IEMD-NFPSO-CNN-ResGRU combined model is constructed to perform fault diagnosis. The model has obvious advantages in cable fault diagnosis.

Description

Technical Field

[0001] The invention relates to a cable insulation fault diagnosis technology, and in particular to a cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU. Background Art

[0002] At present, when an accident occurs in a cable, it will pose a great threat to the stability and security of the power grid. In conventional power grids, small current grounding is a common power distribution method, but practice has proved that when a single-phase short circuit occurs in the power grid, the voltage of the three-phase line is still symmetrical; it will not cause any interference to the power system, and it can work continuously for two hours, but if the grounding current exceeds the prescribed limit, intermittent arcs are likely to occur, causing arc overvoltage and other faults, such as three-phase short circuits. In this case, the action of the relay protection device will cause large-scale power outages, seriously affecting production. Existing offline detection technologies mainly include offline detection and on-site monitoring. The offline detection method can only be performed when a power outage occurs in the power grid, so it is not suitable for power system detection.

[0003] At present, many researchers at home and abroad have adopted the method of real-time monitoring to extract the transient traveling wave information of the fault in the power system and use it for the fault diagnosis of the power system. Some have analyzed the zero-sequence current and integrated the wavelet packet method with information entropy to diagnose its fault. Some have detected various types of power lines in the power system, and used the method of combining wavelet analysis with neural networks to detect various types of power systems in the power system. When using BP neural network for fault detection, there are problems such as slow speed and easy to fall into local extreme values. Some have taken the power system in the power system as the object, the wavelet packet energy entropy of the power system as the feature vector, and adopted the method of combining PSO-BP neural network to realize the online monitoring of the mine power system. Some have used the wavelet filtering method to filter the vibration signal of the fan gearbox and improved it. It is suggested to use incremental convolutional neural network to detect cables online to achieve early fault identification of cables.

[0004] The present invention adopts an S-transformation method to extract the time-frequency domain characteristics of multi-domain signals, adopts a PCA method to perform dimensionality reduction processing on the multi-domain signals, and converts them into input vectors of an incremental convolutional neural network. Summary of the invention

[0005] The main purpose of the present invention is to provide a cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU, which solves the problem of low precision and accuracy of XLPE cable fault diagnosis method.

[0006] The technical solution adopted by the present invention is: a cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU, comprising:

[0007] Generate cable fault simulation data, randomly divide them into training samples and test samples in a ratio of 6:4, and perform preprocessing;

[0008] Initialize the parameters of IEMD, NFPSO, CNN, and ResGRU;

[0009] The improved empirical mode decomposition method IEMD is used to decompose the training samples and test samples;

[0010] A pre-training process is performed for each particle in each cycle. During the pre-training process, the optimal parameter values ​​of the particles are used as the hyperparameters of the model. The number of pre-training is K, and the output is the diagnosis result of the model.

[0011] The diagnostic error calculation formula of the model is set as the fitness value function of the NFPSO algorithm, and the global historical optimal parameters are gradually updated; in each iteration, the fitness value of the particle is calculated according to the fitness value function, and the fitness value is compared with the single historical optimal value of the particle and the global historical optimal value of the population. If the fitness value is better, the historical optimal value of the single particle and the global historical optimal value of the population are updated;

[0012] Determine whether the maximum number of iterations T has been reached. If not, continue the iteration process. If it has reached T times, output the global historical best value, that is, the optimal hyperparameters of CNN and ResGRU.

[0013] The generated CNN and ResGRU with optimal hyperparameters are used for fault diagnosis.

[0014] Advantages of the present invention:

[0015] The present invention proposes a cable fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU. Firstly, the cable fault sample data is decomposed by the IEMD algorithm to obtain the IMF components, and then the constructed CNN-ResGRU model is optimized for hyperparameters using the neighborhood-based fuzzy particle swarm algorithm (NFPSO). The optimized CNN-ResGRU model is used for fault diagnosis, which has obvious advantages in cable fault diagnosis.

[0016] 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

[0017] 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.

[0018] Figure 1 This is the structural diagram of 1D-CNN;

[0019] Figure 2 It is a ring topology graph with 2 neighborhood radii;

[0020] Figure 3 is the membership function of the difference d between two consecutive iterations of f(gbest) of the present invention;

[0021] Figure 4 is the membership function of the inertia weight w of the present invention;

[0022] Figure 5 is the membership function of the learning factor c1 of the present invention;

[0023] Figure 6 is the membership function of the learning factor c2 of the present invention;

[0024] Figure 7 It is the CNN-ResGRU model architecture of the present invention;

[0025] Figure 8 It is the flow chart of the IEMD-NFPSO-CNN-ResGRU model of the present invention;

[0026] Fig. 9 It is the CNN-ResGRU model result of the present invention;

[0027] Fig.10 is a comparison curve of the model loss rate of the present invention;

[0028] Fig.11 It is a comparison curve of the model accuracy of the present invention. DETAILED DESCRIPTION

[0029] 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.

[0030] Improved empirical mode decomposition method:

[0031] The purpose of EMD is to decompose complex non-stationary signals into a finite number of IMFs. The IMFs must satisfy the following conditions:

[0032] (1) The number of extreme values ​​and transitive zeros of the function must be equal to or differ by at most 1;

[0033] (2) The average value of the upper envelope formed by the local maximum point of the function and the lower envelope formed by the local minimum point is zero.

[0034] According to the maximum and minimum values ​​provided by the original non-stationary signal, cubic spline interpolation is performed to obtain the upper and lower envelopes in the EMD processing. However, the two endpoints of the signal may not be extreme points, which will produce fitting errors. As the decomposition process progresses, the error will gradually spread to the internal signal, causing the waveform of the IMF component to be severely distorted. This is called the end effect of EMD. In order to maintain the authenticity of the data, the present invention proposes an improved empirical mode decomposition method-IEMD.

[0035] This method uses the mirror extension method to improve EMD. The mirror extension method is based on the characteristics of mirror symmetric mapping. In order to minimize side effects, the mirror is placed in a position to ensure that the signal has symmetrical extreme values. The specific steps are as follows:

[0036] (1) Mirror extension preprocessing: Identify the extreme points (maximum and minimum) at both ends of the original signal. Determine the mirror position based on the signal distribution characteristics. Extend the two ends of the signal through mirror symmetry to form a continuous symmetrical signal. The length of the extended signal is usually twice that of the original signal, which is used to reduce the dependence on the end data during the decomposition process.

[0037] (2) Constructing envelopes (upper envelope and lower envelope): Using local maximum points and local minimum points, construct the upper envelope and lower envelope respectively through interpolation methods (such as spline interpolation). Ensure that the envelope can tightly wrap the signal while maintaining smoothness.

[0038] (3) Calculate the mean of the signal: Take the average of the upper envelope and the lower envelope to obtain the envelope mean curve m(t)m(t)m(t). Subtract the mean curve from the signal to obtain a candidate intrinsic mode function (IMF).

[0039] (4) Optimize the IMF screening (improve the screening process): Improve the stopping criteria for the candidate IMF: Check whether the monotonicity or zero crossing point of the candidate IMF meets the IMF definition; introduce adaptive thresholds or improved criteria to enhance the accuracy and stability of the decomposition. If the IMF conditions are not met, repeat steps (2) and (3) until an IMF that meets the conditions is screened out.

[0040] (5) Remove the current IMF and recursively decompose: Remove the obtained IMF from the original signal to obtain the residual signal. Repeat the above steps for the residual signal and continue to decompose the next layer of IMF until the residual signal tends to be monotonic or meets the termination condition.

[0041] (6) Handling end effects: During the entire decomposition process, the mirrored extended signal is used to ensure the integrity of the boundary data and avoid end errors. When the decomposition is completed, the effective part of the extended signal is intercepted and restored to the original length.

[0042] (7) Post-processing (optional): Perform spectrum or feature analysis on all IMFs and perform further processing based on application requirements. The remaining residual signal usually represents a trend term and can be directly used as part of the decomposition result.

[0043] Neighborhood-based fuzzy particle swarm optimization algorithm:

[0044] PSO is a population-based stochastic meta-heuristic optimization algorithm inspired by the ability of certain animal species, such as flocks of birds or schools of fish, to locate desired positions in a specified area. The swarm consists of a set of particles, each of which represents a potential solution vector for the optimization problem. The position of each particle is dynamically updated based on its previous position and the new velocity, which is evaluated by learning from its own experience and other particles in the swarm. The current position and velocity vector of a particle at iteration are denoted by and , respectively. The most famous position in the individual history is the best position found so far in the particle's search space. The position of the global best particle is denoted by . During the search, the velocity and position vectors of the particles are updated by the following equations:

[0045] V i g+1 =wV i g +c1r1(pbest i -X i g )+c2r2(gbest-X i g ) (1)

[0046] X i g+1 =X i g +V i g+1 (2)

[0047] X min ≤X i g+1 ≤X max ,i=1,2,...,N (3)

[0048] Where i is the number of particles, N is the group size, w is the inertia weight, c1 and c2 are acceleration coefficients, r1 and r2 are random numbers between 0 and 1, and X min With X max They represent the lower and upper bounds of the search space respectively.

[0049] The standard PSO algorithm suffers from early convergence problems when solving highly multimodal problems. In the present invention, a neighborhood-based fuzzy PSO algorithm is proposed to trade off between exploration (global search capability) and exploitation (local search capability).

[0050] Neighborhood Learning Strategy

[0051] The neighborhood of a particle trapped in a suboptimal value can contain the global optimum. In this case, searching the neighborhood of the particle is very useful for locating the global optimum. When the suboptimal value is close to the global optimum, the neighborhood search strategy can effectively solve the multimodal problem.

[0052] The present invention proposes a neighborhood-based PSO algorithm learning strategy, which considers the local best particle in the particle neighborhood in the velocity update rule instead of the global best particle to avoid falling into the local optimum. Here, all particles are members of a neighborhood constructed by a certain topological design, and the learning of particles is limited to this neighborhood.

[0053] To implement the neighborhood, a ring topology is used for simplicity and the length of the neighborhood is set to K, as complex topologies increase the runtime complexity of the algorithm. The two neighboring particles on either side of a particle in the ring are identified as the neighbors of that particle. Figure 2 shows a ring topology with 2 particles in the swarm. Let P i (i=1,2,…,N) is the i-th particle in the group, where N is the size of the group. Consider N+1=1 and N+2=2, where N is the number of particles in the group. The N particles are arranged in a ring topology according to their indices, such that P N and P2 are two neighbors of P1. Each particle P i The k-neighborhood radius of particle P i-k , P k ……,P i+k , where k is an integer, 0≤k≤N-1. Experimental results show that the performance of the algorithm is not affected by the parameter k, and its performance is basically the same for different k values. Figure 3 shows the k-neighborhood radius when k = 2. If the index of a particle is i, then the neighborhood of the i-th particle includes particles with indices i-2, i-1, i+1, and i+2, as shown in Figure 2 shown.

[0054] Let lbest be the position of the local best particle, that is, the position of the best particle in the particle neighborhood. Then (1) becomes as follows:

[0055]

[0056] Now, the particles learn from themselves at the same time. According to (4), each particle will move towards its own best position and the best position of its neighborhood, rather than towards the global best position as in (1).

[0057] Adaptive fuzzy scheme

[0058] In order to prevent the particle swarm algorithm from reaching local optimum during operation, the present invention considers the fuzzy adaptability of the particle swarm algorithm parameters and adopts fuzzy IF / THEN rules to adaptively control the inertia weight w and learning factors c1 and c2 in the PSO algorithm (1) to improve the algorithm efficiency and search accuracy.

[0059] Let f(gbest g ) represents the best fitness function value at iteration g, and d represents the difference between f(gbest) at two consecutive iterations. At iteration g, the evaluation of d (for the minimization problem) is as follows:

[0060] d g =f(gbest g-1 )-f(gbestg)≥0 (5)

[0061] When d is large, especially at the beginning of the search process, w and c1 should be increased to thoroughly explore the search space to find a global solution, and c2 should be reduced to a smaller value to avoid premature termination of the search. As the value of d decreases, the particle swarm converges to an optimal solution, and the area around the optimal solution needs to be thoroughly searched. Therefore, w and c1 are set to smaller values, and c2 is increased to a larger value to increase the convergence speed. When d is medium, w, c1, and c2 are kept at medium values. On this basis, a fuzzy system with d as the input variable, w, c1, and c2 as the output variables, and the inertia weight and learning factor are adjusted is proposed.

[0062] To fuzzify each input and output, use Figure 3 Membership functions of d, w, c1, and c2 are shown.

[0063] Here, Small, Medium, and Large are the linguistic values ​​of input and output defined by fuzzy sets. The following Mamdani-type fuzzy rules are proposed to express conditional statements involving fuzzy logic. Rule 1: If d is small,

[0064] Then w is small, c1 is small, and c2 is large. Rule 2: If d is medium,

[0065] Then w is medium, c1 is medium, and c2 is medium. Rule 3: If d is large,

[0066] Then w is large, c1 is large, and c2 is small.

[0067] Each rule describes the mapping from input space to output space. In order to obtain a clear value, the output must be defuzzified. The present invention uses the centroid method to defuzzify each output.

[0068] Let the maximum number of iterations g max It is divided into q partitions, represented by qi,i=0,1,…,q-1, and the partition qi consists of iterations. Therefore:

[0069]

[0070] After ni iterations of partition qi, Figure 3 The values ​​of d' and d1 in n are modified once. Obviously, i After iteration, n i The new d is generated by (5). The values ​​of d' and d1 are updated as follows:

[0071]

[0072] d' and d1 are not applicable for the first n0 iterations. Therefore, the present invention assumes the subsequent part of fuzzy rule 3 by choosing larger constants for w and c1 (w=0.8 and c1=2.5) and a smaller constant for c2 (c2=0.5).

[0073] Cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU:

[0074] For deep learning model structures, changes in hyperparameters have a more significant impact on the model, and reasonable parameter adjustment is also a very important process in model training. The particle swarm algorithm has a good effect on the optimization of nonlinear problems, so before officially entering the training process, the improved particle swarm algorithm is used to optimize the hyperparameters of the algorithm model.

[0075] CNN-ResGRU model:

[0076] Convolutional neural network (CNN) can fully explore the relationship between various local features, and ResGRU can memorize the mined temporal dynamic features, making it easier for the model to capture the correlation and dependency features between insulation fault data. The present invention proposes a fault diagnosis model (CNN-RerGRU) combining CNN and ResGRU. Based on the end-to-end machine learning model, the fault features and rules contained in the insulation fault data are effectively mined and extracted, which simplifies the diagnosis process and improves the recognition accuracy.

[0077] The CNN-ResGRU model structure is as follows Figure 7As shown in the figure, it includes 1-DCNN module, ResGRU module and fully connected module. Specifically, it mainly consists of input layer, convolution layer, pooling layer, ResGRU layer and output layer. The information of feature information is extracted through the convolution layer, and then the features are reduced through the pooling layer. After that, the ResGRU module is used to memorize and digest the rules contained in the previous features, and finally the classification is performed through the fully connected layer.

[0078] The DCNN module outputs two-dimensional features, including time and space (channel) dimensions. Specifically, the function of ResGRU is explained as follows. The first GRU stores and mines the spatial dimension information to return data with a complete time dimension. The output result is converted into the second GRU to achieve the storage and mining of time dimension information. However, the result of this processing is that the input and output dimensions of the residual block do not match. Therefore, in the identity mapping part, dimension matching needs to be completed. The processing method of ResNet is to use a 1×1 convolution kernel to achieve channel dimensionality reduction. However, due to the change in the time scale of the output data, simple convolution cannot be achieved. Therefore, the GRU module is used in identity mapping to achieve dimension matching. Compared with ordinary GRU units, ResGRU not only has independent residual learning functions, but also proposes features of different time and space scales from the original information through multiple connected pipelines.

[0079] Output layer. It is essentially a fully connected layer, and its function is classification. Therefore, softmax is selected as the activation function. In this layer, the model calculates the probability of each input sample and then obtains a new expression (y predict ).

[0080] y predict (i) = f(L = l i |S6;(W,b)) (9)

[0081] Where f(.) represents the softmax activation function; l i It represents the probability that the input sample belongs to the i-th category; W represents the weight and b represents the bias value.

[0082] For model training, random numbers are usually used to initialize the weight matrix and bias values ​​to ensure that each parameter is not repeated and the difference is not large. Use he_normal to achieve this work.

[0083] The data is made to have a good constant variance when it is input to the first convolutional layer. In addition, L2 regularization is selected in the output layer to accelerate the convergence speed of the network and prevent the network from overfitting. In addition, cross entropy is used to process the output probability model.

[0084] IEMD-NFPSO-CNN-ResGRU model:

[0085] The cable fault diagnosis problem can be attributed to the time series decomposition problem. In order to further improve the accuracy of fault diagnosis, this paper decomposes the fault data based on the proposed IEMD, optimizes the hyperparameters of the CNN-ResGRU model based on the NFPSO algorithm proposed in Section 3, and constructs a cable insulation fault diagnosis model based on IEMD-NFPSO-CNN-ResGRU (IEMD-NFPSO-CNN-ResGRU). The specific implementation process is as follows: Figure 8 The overall implementation steps are as follows:

[0086] (1) Generate cable fault simulation data, randomly divide them into training samples and test samples in a ratio of 6:4, and perform preprocessing.

[0087] (2) Initialize the relevant parameters of IEMD, NFPSO, CNN, and ResGRU.

[0088] (3) Use IEMD to decompose training samples and test samples;

[0089] (4) Perform a pre-training process for each particle in each cycle. In the pre-training process, the optimal parameter values ​​of the particles are used as the hyperparameters of the model. The number of pre-training is K, and the output is the diagnosis result of the model.

[0090] (5) The diagnostic error calculation formula of the model is set as the fitness value function of the NFPSO algorithm, and the global historical optimal parameters are gradually updated. In each iteration, the fitness value of the particle is calculated according to the fitness value function, and the fitness value is compared with the single historical optimal value of the particle and the global historical optimal value of the population. If the fitness value is better, the historical optimal value of the single particle and the global historical optimal value of the population are updated.

[0091] (6) Determine whether the maximum number of iterations T has been reached. If not, continue the iterative process; if it has reached T times, output the global historical best value, that is, the optimal parameters of CNN and ResGRU.

[0092] (7) Use the CNN-ResGRU with optimal hyperparameters generated in (6) for fault diagnosis.

[0093] Experimental results and comparative analysis

[0094] NFPSO algorithm performance verification:

[0095] In order to verify the effectiveness of the NFPSO proposed in this paper, four well-known benchmark functions in this paper are firstly experimented to determine the relevant optimal parameters. In particular, the performance of PSO with different initialization methods and different inertia weights is compared and studied. Note that all the test functions are shown in Table 1, including their expressions, dimensions, search spaces, allowed search ranges and global optimal values.

[0096] Table 1 Benchmark functions

[0097]

[0098] In order to make a fair comparison with several state-of-the-art PSO variants (including NIPSO, QPSO, FRBPSO, TS-PSO and E-μPSO), the present invention runs 30 times on each PSO on each test function, and each run is iterated 1000 times. The experimental results are shown in Table 2.

[0099] From Table 2, we can find that MPSO has better optimization accuracy than the other PSOs and achieves the best results on three of the four test functions. Although it is not optimal on the other one function, the optimization accuracy is still within an acceptable range.

[0100] Table 2 Comparison of experimental results of NFPSO and other PSOs

[0101]

[0102] Cable Fault Diagnosis Example Analysis

[0103] Test data:

[0104] The actual cable fault simulation was carried out and the fault current signal on the cable line was collected. The collected data included four types of cable fault signals, each type of fault contained 5000 samples, and the total number of samples was 20000. The sampling of various signal samples is shown in Table 3.

[0105] Table 3 Number of samples of various signals

[0106]

[0107]

[0108] Evaluation indicators:

[0109] The present invention uses accuracy as the model performance evaluation index. The specific calculation formula is as follows:

[0110]

[0111] Among them, TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

[0112] Model parameter settings:

[0113] In the MPSO-1DCNN-BiLSTM model, the number of iterations is set to 500, and the parameter settings of the MSPO algorithm are consistent with the above. The 1DCNN consists of 1 convolution layer (filter size is 3*3) and 1 pooling layer (MaxPooling), and the activation function is RELU; the activation function in BiLSTM is Tanh, and the range of n_hidden is set to 256 to 272. The batch size is 128, the learning rate decay coefficient of the two networks is set from 0.045 to 0.055, the weight decay coefficient is set to 0.00085 to 0.00095, and the optimizer is Adam.

[0114] IMF component selection:

[0115] In the IEMD algorithm, the selection of the number of IMF components after decomposition is the most important, and the correlation between the first few IMF components and the original signal is relatively large. The present invention performs IEMD decomposition on the fault data. After IEMD decomposition, the correlation coefficient between each IMF component and the original signal is calculated, as shown in Table 4. It can be seen that the contribution rate of the first 8 components of the 13 IMFs reaches 92.23%, which contains the main information of the original data. Therefore, the present invention uses them to characterize the original data.

[0116] Table 4 Correlation coefficients between each IMF component and the original signal

[0117]

[0118]

[0119] Comparison between CNN-ResGRU, NFPSO-CNN-ResGRU and IEMD-NFPSO-CNN-ResGRU

[0120] In order to verify the effectiveness of IEMD decomposition and the effectiveness of NFPSO algorithm optimization, the present invention compares and analyzes CNN-ResGRU, NFPSO-CNN-ResGRU and IEMD-NFPSO-CNN-ResGRU. The fault diagnosis accuracy change curves of the three models on training data and test data are shown in the figure. Fig. 9As shown. It can be seen that the IEMD-NFPSO-CNN-ResGRU model has better diagnostic results than the other two models. On the one hand, this is because the IEMD method used in the IEMD-NFPSO-CNN-ResGRU model decomposes the original fault data, and the decomposed IMF components with more feature information are used as the input of NFPSO-CNN-ResGRU; on the other hand, the NFPSO algorithm can search for better hyperparameters during the hyperparameter optimization process, which makes the CNN-ResGRU have better diagnostic accuracy when performing fault diagnosis.

[0121] The fault diagnosis accuracy of the three models on the test data is shown in Table 5.

[0122] Table 5 Comparison of fault diagnosis accuracy of three models

[0123]

[0124] As can be seen from Table 5, the CNN-ResGRU model has a strong feature extraction capability, but because the relevant hyperparameters are set according to experience, its accuracy is the lowest, only 94.52%; the NFPSO-CNN-ResGRU model uses the PSO algorithm for hyperparameter optimization, and its accuracy reaches 97.23%, which is 2.71% higher than the CNN-ResGRU model; the IEMD-NFPSO-CNN-ResGRU model uses the IEMD and NFPSO algorithms, which further improves its accuracy to 98.61%.

[0125] Comparison with existing models:

[0126] The IEMD-NFPSO-CNN-ResGRU model is compared and analyzed with the existing models, and EEMD-BAS-CNN, PSO-SVM, CWT, and WT-CNN-insformer are used as comparison models for experiments. The accuracy and error change curves of the six models are shown in the figure. Fig.10 , Fig.11 As shown, the fault diagnosis results are compared in Table 6.

[0127] Table 6 Comparison of model fault diagnosis results

[0128]

[0129] from Fig.10 , Fig.11 ,It can be seen from Table 6 that the IEMD-NFPSO-CNN-ResGRU model has the best ,diagnosis results compared with other models.

[0130] This is mainly because: the model uses the IEMD algorithm to decompose fault data, and uses the decomposed IMF component data for fault diagnosis to obtain better results; the NFPSO algorithm is used to optimize the CNN-ResGRU model to obtain better hyperparameters. This fully demonstrates that the IEMD-NFPSO-CNN-ResGRU model has obvious advantages in cable fault diagnosis because it combines the advantages of IEMD, NFPSO, CNN, and ResGRU.

[0131] Aiming at the problem of low precision and accuracy of traditional XLPE cable fault diagnosis methods, the present invention proposes a cable insulation fault diagnosis based on IEMD-NFPSO-CNN-ResLSTM.

[0132] Firstly, an improved empirical mode decomposition (IEMD) method is used to decompose the cable fault sample data to obtain the intrinsic mode function (IMF);

[0133] Then, a fusion model based on convolutional neural network (CNN) and residual gated recurrent unit (ResGRU) was designed for pre-training;

[0134] Finally, a neighborhood-based fuzzy particle swarm optimization algorithm (NFPSO) was designed to optimize the hyperparameters of the CNN-ResGRU model, and an IEMD-NFPSO-CNN-ResGRU combined model was constructed for fault diagnosis.

[0135] Experimental results show that the NFPSO algorithm has better optimization ability and faster convergence speed than the existing PSO improved algorithm. The classification accuracy of the IEMD-NFPSO-CNN-ResGRU model reaches 98.61%, which is significantly better than other existing models.

[0136] In the experiment, on the one hand, the NFPSO algorithm is compared with the NIPSO, QPSO, FRBPSO, TS-PSO and E-μPSO algorithms. The results show that the algorithm has better performance in the optimization of the benchmark function;

[0137] On the other hand, the IEMD-NFPSO-CNN-ResGRU model was compared with other existing models. The results show that this model fully combines the advantages of IEMD, NFPSO, CNN and ResGRU, and has obvious advantages in cable fault diagnosis.

[0138] 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 diagnosis model based on IEMD-NFPSO-CNN-ResGRU, characterized in that: include: Generate cable fault simulation data, randomly divide them into training samples and test samples in a ratio of 6:4, and perform preprocessing; Initialize the parameters of IEMD, NFPSO, CNN, and ResGRU; The improved empirical mode decomposition method IEMD is used to decompose the training samples and test samples; A pre-training process is performed for each particle in each cycle. During the pre-training process, the optimal parameter values ​​of the particles are used as the hyperparameters of the model. The number of pre-training is K, and the output is the diagnosis result of the model. The diagnostic error calculation formula of the model is set as the fitness value function of the NFPSO algorithm, and the global historical optimal parameters are gradually updated; in each iteration, the fitness value of the particle is calculated according to the fitness value function, and the fitness value is compared with the single historical optimal value of the particle and the global historical optimal value of the population. If the fitness value is better, the historical optimal value of the single particle and the global historical optimal value of the population are updated; Determine whether the maximum number of iterations T has been reached. If not, continue the iteration process. If it has reached T times, output the global historical best value, that is, the optimal hyperparameters of CNN and ResGRU. The generated CNN and ResGRU with optimal hyperparameters are used for fault diagnosis.

Images