CN-121980364-A - Cable fault diagnosis method and device based on neural network model

CN121980364ACN 121980364 ACN121980364 ACN 121980364ACN-121980364-A

Abstract

The application discloses a cable fault diagnosis method and device based on a neural network model, and relates to the field of power system equipment state monitoring and fault diagnosis, wherein the method comprises the steps of collecting and preprocessing time domain signals of a cable in different states, adopting a PSO algorithm to improve CAOA algorithm to obtain a CAPSO algorithm, and adopting the CAPSO algorithm to optimize TVFEMD algorithm parameters; extracting characteristics of the optimal IMF components and performing characteristic dimension reduction processing, and training a neural network model based on the characteristics after the characteristic dimension reduction processing so as to realize the identification of the cable fault type. The application can realize the accurate diagnosis of cable faults.

Inventors

LIAO JIAN
TANG LONG
Mi Luqi
KE XI
ZHANG MENG
HUANG LI
YANG PENGYU
CHEN PEILIN
ZHU SHA
LONG TAO
XING LEI

Assignees

武汉三相电力科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The cable fault diagnosis method based on the neural network model is characterized by comprising the following steps of: Collecting and preprocessing time domain signals of the cable in different states, adopting a PSO algorithm to improve CAOA algorithm to obtain a CAPSO algorithm, and adopting the CAPSO algorithm to optimize parameters of the TVFEMD algorithm; Decomposing the time domain signal based on the TVFEMD algorithm after parameter optimization to obtain a plurality of IMF components and determining an optimal IMF component; Extracting the characteristics of the optimal IMF component, performing characteristic dimension reduction processing, and training a neural network model based on the characteristics after the characteristic dimension reduction processing so as to realize the identification of the cable fault type.
2. The cable fault diagnosis method based on the neural network model as claimed in claim 1, wherein the time domain signals of the cable in different states are collected and preprocessed, specifically comprising: a high-frequency current sensor and an ultrahigh-frequency partial discharge sensor are deployed at a set key position of a cable line so as to acquire time domain signals of the cable in normal operation and different fault states, wherein the fault states comprise high-resistance flashover, low-resistance short circuit, insulation aging and wire breakage; and carrying out data preprocessing on the acquired time domain signals based on median filtering, db4 wavelet basis and normalization processing mode.
3. The cable fault diagnosis method based on the neural network model as claimed in claim 1, wherein the improving CAOA algorithm of the PSO algorithm is adopted to obtain a CAPSO algorithm, and the optimizing TVFEMD algorithm parameters are adopted to optimize the CAPSO algorithm, specifically comprising: Determining parameters to be optimized of a TVFEMD algorithm and a search range of the parameters to be optimized, wherein the parameters to be optimized of the TVFEMD algorithm comprise a filter bandwidth, an iteration stop threshold value and an initial decomposition layer number; Carrying out initialization setting of population scale and core parameters of CAOA algorithm, and setting inertia weight of PSO algorithm into linear decreasing mode; Constructing a multi-target weighted combination fitness function, and determining indexes for evaluating global search states, wherein the indexes comprise global optimal solution stability, population diversity and fitness convergence speed; Repeatedly executing the iterative operation until the iterative termination condition is met, and outputting TVFEMD the optimal parameter combination after the algorithm optimization; The iteration operation is that a global exploration stage is entered, population updating in the global stage is carried out, when any index for evaluating the global search state is met, a local refinement stage is entered, population updating in the local refinement stage is carried out, new fitness after individual updating is calculated after each iteration and is compared with the historical optimal fitness of the individual, so that the optimal parameters of the individual and the optimal fitness of the individual are updated, and the updated optimal fitness of the individual and the global optimal fitness are compared, so that the optimal parameters of the global and the optimal fitness of the global are updated.
4. A cable fault diagnosis method based on a neural network model as claimed in claim 3, wherein: The multi-objective weighted combination fitness function specifically comprises the following steps: Wherein, the Representing a multi-objective weighted combination fitness function, The spectral entropy is represented by the value of the spectral coefficients, Represents the modal aliasing suppression index, Representing the signal-to-noise ratio of the fault signature, Representing the effective IMF component duty cycle; The said The concrete steps are as follows: Wherein, the The number of frequency sampling points representing the optimal IMF component, Represent the first Power spectral density values for the individual frequency points; The said The concrete steps are as follows: Wherein, the And Representing two adjacent IMF components; The said The concrete steps are as follows: Wherein, the Representation of The number of IMF components is determined to be effective.
5. A cable fault diagnosis method based on a neural network model as claimed in claim 3, wherein: The global optimal solution stability is specifically expressed as follows: Wherein, the Represents the stability of the globally optimal solution, Representing the number of current iterations and, Represent the first The global optimum fitness is iterated a number of times, Represent the first The global optimum fitness is iterated a number of times, Representing a stable iteration number; The population diversity is specifically expressed as: Wherein, the The diversity of the population is represented by the number of the species, Represent the first Individual first The dimensional parameters of the object are defined by the dimensions, Represent the first The population mean value of the dimensional parameters, Represent the first The search range of the dimension parameter is defined, Representing the total number of individuals; The adaptability convergence speed is specifically expressed as: Wherein, the Indicating the convergence speed of the fitness.
6. A cable fault diagnosis method based on a neural network model as claimed in claim 3, wherein said entering a global exploration phase, and performing population update in the global phase, specifically comprises: entering a global exploration stage, leading CAOA algorithm to traverse a wide parameter space, and weakening a local search function by PSO algorithm; Leading the population to be close to a high-quality parameter area through a leader selection mechanism of CAOA algorithm, wherein the selection standard of the leader is that the smaller the fitness value is, the higher the probability is selected, and the probability is calculated in the following way: Wherein, the Represent the first The probability that an individual is selected as the leader, Representing a multi-objective weighted combination fitness function, Representing the total number of individuals; the crocodile position updating mode is determined, specifically: Wherein, the Represent the first First iteration The position of the individual crocodiles is, Represent the first The position of the leader of the next iteration, Representing the strength of the attraction of the leader, Representing the amplitude of the random disturbance, Representing a random number; the energy consumption mechanism of crocodile hunting is simulated, the more the individual parameter moves the distance, the more the energy consumption, the energy attenuation formula is: Wherein, the The energy consumption rate is indicated by the energy consumption rate, Represent the first First iteration Residual energy of crocodile when Less than or equal to 0, re-randomly generating within the parameter search range And reset At this time, the PSO algorithm only performs weakened local adjustment, and the speed update formula is: Wherein, the Representing a linear decrease in the inertial weight, Representing the number of current iterations and, Represent the first First iteration The speed of the individual crocodiles is such that, 、 The learning factor is represented as such, Represent the first The historical optimal location of the individual(s), The number of iterations of the maximum is indicated, Represent the first The global optimum position for the next iteration, 、 Representing a random number.
7. The cable fault diagnosis method based on a neural network model as claimed in claim 6, wherein the entering a local refinement stage, and the population updating in the local refinement stage, specifically comprises: Entering a local refinement stage, wherein the PSO algorithm dominates the accurate search at the moment, and CAOA algorithm only reserves a basic mechanism to prevent population stagnation; the PSO algorithm restores the complete local refinement capability, the speed updating formula is kept unchanged, the learning factor is increased, the speed is subjected to boundary control, and the crocodile position updating mode is adjusted by the following steps: 。
8. The method for diagnosing a cable fault based on a neural network model as claimed in claim 4, wherein said TVFEMD algorithm based on the parameter optimization decomposes the time domain signal to obtain a plurality of IMF components and determines an optimal IMF component, specifically comprising: loading the optimized optimal parameter combination in TVFEMD algorithm; Using the time domain signal as the input of TVFEMD algorithm, separating to obtain multiple IMF components, and obtaining A significant IMF component; Using spectral entropy Signal to noise ratio with fault signature And directly adding an equalization optimization strategy of the minimum value, and determining to obtain the optimal IMF component.
9. The cable fault diagnosis method based on the neural network model as claimed in claim 1, wherein the extracting the characteristics of the optimal IMF component and performing the characteristic dimension reduction processing, training the neural network model based on the characteristics after the characteristic dimension reduction processing to realize the identification of the cable fault type, specifically comprises: extracting time domain features, frequency domain features and time-frequency domain features of the optimal IMF components, and performing feature dimension reduction by analyzing WM-PCA through a weighted principal component; Dividing the feature after dimension reduction into a training set, a testing set and a verification set according to a proportion, so as to train and verify the constructed neural network model, and evaluating the accuracy to obtain a trained neural network model capable of carrying out cable fault type recognition; And acquiring time domain signals of the cable to be subjected to fault diagnosis, processing the acquired time domain signals based on the cable fault diagnosis method to obtain characteristics after feature dimension reduction processing, inputting the characteristics into a trained neural network model, and carrying out cable fault type identification.
10. A cable fault diagnosis apparatus based on a neural network model, characterized in that the cable fault diagnosis apparatus based on a neural network model includes: The optimizing module is used for collecting and preprocessing time domain signals of the cable in different states, adopting a PSO algorithm to improve CAOA algorithm to obtain a CAPSO algorithm, and adopting the CAPSO algorithm to optimize parameters of the TVFEMD algorithm; The decomposition module is used for decomposing the time domain signal based on the TVFEMD algorithm after parameter optimization to obtain a plurality of IMF components and determining an optimal IMF component; The training module is used for extracting the characteristics of the optimal IMF component and performing characteristic dimension reduction processing, and training the neural network model based on the characteristics after the characteristic dimension reduction processing so as to realize the identification of the cable fault type.

Description

Cable fault diagnosis method and device based on neural network model Technical Field The application relates to the field of power system equipment state monitoring and fault diagnosis, in particular to a cable fault diagnosis method and device based on a neural network model. Background The power cable is used as a core component of the power transmission network and bears the key task of electric energy transmission, and the operation reliability of the power cable is directly related to the power supply continuity of the power grid and the safety and stability of the power system. In recent years, a new path is provided for cable fault diagnosis by fusion of a signal processing technology and a deep learning technology, a time-varying filtering empirical mode decomposition algorithm (TVFEMD) is led into cable fault signal processing due to the fact that the time-varying filtering empirical mode decomposition algorithm can adapt to non-stationary signal decomposition requirements, core parameters such as filtering bandwidth, iteration threshold and the like of the algorithm are set depending on manual experience, mode aliasing or feature loss is easy to occur, a Convolutional Neural Network (CNN) has strong feature extraction capability, but effective fault features are difficult to focus when original fault signals are processed directly, and diagnosis accuracy is limited under complex noise interference. Specifically, the existing cable fault diagnosis method has the following problems: (1) On one hand, the global search strategy of the crocodile volt-ampere optimization algorithm (CAOA) can cover a wide parameter space, effectively traverses a potential optimal solution area, but has insufficient excavation depth on a local area around an approximate optimal solution, is easy to generate search concussion near an optimal interval and is difficult to accurately converge to the global optimal solution, on the other hand, the local refinement capability of CAOA depends on initial parameter setting, lacks an adaptive mechanism for dynamically adjusting search precision, and has insufficient flexibility of the search strategy of the single algorithm, and cannot simultaneously consider 'wide global exploration' and 'high-precision local convergence', so that upper limits exist on result stability and precision when the crocodile volt-ampere optimization algorithm is independently optimized; (2) The TVFEMD parameter optimization mode is low in efficiency, high-quality signal decomposition is difficult to realize, two main modes exist in the TVFEMD parameter configuration, obvious defects exist in the TVFEMD parameter configuration, firstly, the TVFEMD parameter optimization mode is manually set, the manual setting is completely dependent on engineering experience of operators, the subjectivity is strong, signal characteristics of cables with different fault types and voltage levels are extremely different, single experience parameters cannot adapt to multiple scenes and are easy to cause modal aliasing or insufficient decomposition, and secondly, a traditional single group intelligent algorithm is adopted, or a local optimal solution is easy to fall into, global optimal parameter combinations are difficult to search, or a global exploration range is limited, a complete parameter space cannot be covered, or the local refinement precision is insufficient, and the similar optimal parameters are difficult to further converge after being found; (3) The cable fault diagnosis model is low in precision, the key reasons are from linkage influence of a preamble link, the TVFEMD parameter optimization inefficiency causes that the decomposed IMF component contains a large amount of redundant noise or key feature loss, the feature quality of an input classification model is low, the model is difficult to learn fault features with strong distinguishability, classification confusion is easy to occur, in actual working conditions such as multi-branch cables, complex electromagnetic interference and the like, the diagnosis accuracy is greatly reduced, and the core requirement of the intelligent power grid on accurate fault diagnosis identification cannot be met. Disclosure of Invention The application provides a cable fault diagnosis method and device based on a neural network model, which can realize accurate diagnosis of cable faults. In a first aspect, an embodiment of the present application provides a cable fault diagnosis method based on a neural network model, where the cable fault diagnosis method based on the neural network model includes: Collecting and preprocessing time domain signals of the cable in different states, adopting a PSO algorithm to improve CAOA algorithm to obtain a CAPSO algorithm, and adopting the CAPSO algorithm to optimize parameters of the TVFEMD algorithm; Decomposing the time domain signal based on the TVFEMD algorithm after parameter optimization to obtain a plurality of IMF components and d