A method and system for processing abnormal power electronics data

Abstract

The present application provides a method and a system for processing abnormal power electronics data, which relates to the field of power electronics data processing. The processing method includes: acquiring an electronic signal of an electronic component to be processed; converting the electronic signal into a digital signal; analyzing the digital signal by using a GA-FCMNN model trained with power electronics parameter-based fault samples; comparing the result of the analysis with a preset normal result, and if there is an abnormality, processing the electronic component to be processed. The genetic algorithm GA is used for the initial parameter selection of the FCMNN, and through mutation operations and optimal selection, the global optimization of the FCNMNN classifier is achieved. The present application can effectively reduce the blindness and time cost of manually selecting initial parameters, further improve the learning efficiency and intelligent level of the neural network diagnostic classifier, and improve the recognition efficiency of electronic component faults.

Description

Technical Field

[0001] This application belongs to the field of power electronic data processing, and particularly relates to a method and system for processing abnormal power electronic data. Background Art

[0002] With the rapid development of power electronic device technology, power electronic equipment is increasingly widely used in various industries, but the problem of equipment failure has gradually increased. Therefore, carrying out research on the diagnosis of early parametric faults in power electronic circuits (such as the aging of components such as electrolytic capacitors and inductors) is of great significance for ensuring the reliability of circuit operation and being able to predict fault risks before faults occur to avoid further hazards. The Cerebellar Model Neural Network (CMNN) is a locally approximated neural network with a simple structure, superior to the neural network, and a fast convergence speed. However, the storage space of the weight coefficients of CMNN will increase sharply with the increase of the input dimension, and the real-time adjustment ability is not strong. The Fuzzy Cerebellar Model Neural Network (FCMNN) introduces the fuzzy theory and makes up for the weakness of self-adjustment of CMNN through the fuzzification of the input, further improving the accuracy and reliability. However, the network structure parameters of FCMNN are manually selected with large randomness, which easily leads to low training efficiency, resulting in a decline in network performance and affecting the approximation ability of the network. Summary of the Invention

[0003] The main purpose of the embodiments of the present invention is to provide a method and system for processing abnormal power electronic data. Through the solution of the embodiments of the present invention, the recognition efficiency of electronic component faults can be improved.

[0004] In a first aspect, a method for processing abnormal power electronic data is provided, including:

[0005] Obtain the electronic signal of the electronic component to be processed;

[0006] Convert the electronic signal into a digital signal;

[0007] Use the genetic algorithm optimized fuzzy cerebellar neural network GA-FCMNN model trained with power electronic parametric fault samples to analyze the digital signal;

[0008] Compare the result of the analysis with the preset normal result. If there is an abnormality, process the electronic component to be processed.

[0009] In a possible implementation manner, before obtaining the electronic signal of the electronic component to be processed, the processing method further includes:

[0010] Set anti-external circuit interference device.

[0011] In another possible implementation, before using the genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronic parameter type fault samples to analyze the digital signal, the processing method further includes:

[0012] Train the GA-FCMNN model with power electronic parameter type fault samples.

[0013] In yet another possible implementation, the training of the GA-FCMNN model with power electronic parameter type fault samples includes:

[0014] Establish a power electronic parameter type fault sample library and establish a GA-FCMNN model;

[0015] Train the GA-FCMNN model with the power electronic parameter type fault samples.

[0016] In yet another possible implementation, the processing of the electronic component to be processed includes: cutting off the power supply of the electronic component to be processed and / or giving an alarm.

[0017] In a second aspect, a processing system for power electronic abnormal data is provided, and the processing system includes:

[0018] An electronic signal acquisition module, configured to acquire an electronic signal of an electronic component to be processed;

[0019] A conversion module, configured to convert the electronic signal into a digital signal;

[0020] An analysis module, configured to analyze the digital signal using a genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronic parameter type fault samples;

[0021] A processing module, configured to compare the result of the analysis with a preset normal result, and if there is an abnormality, process the electronic component to be processed.

[0022] In a possible implementation, the processing system further includes:

[0023] A setting module, configured to set an anti-external circuit interference device.

[0024] In another possible implementation, the processing system further includes:

[0025] A training module, configured to train the GA-FCMNN model with power electronic parameter type fault samples.

[0026] In yet another possible implementation, the training module includes

[0027] a building unit configured to build a power electronics parameter-based fault sample library and build a GA-FCMNN model;

[0028] a training unit configured to train the GA-FCMNN model with the power electronics parameter-based fault samples.

[0029] In yet another possible implementation, processing the to-be-processed electronic component includes: cutting off the power supply of the to-be-processed electronic component and / or giving an alarm.

[0030] The beneficial effects brought by the technical solution provided by this application are: effectively reducing the blindness and time cost of manually selecting initial parameters, further improving the learning efficiency and intelligent level of the neural network diagnostic classifier, and improving the recognition efficiency of electronic component faults. Description of the Drawings

[0031] In order to more clearly illustrate the technical solutions in the embodiments of this application, the following will briefly introduce the drawings required for the description of the embodiments of this application.

[0032] Figure 1 is a flowchart of a method for processing power electronics abnormal data provided by an embodiment of the present invention;

[0033] Figure 2 is a flowchart of a method for processing power electronics abnormal data provided by another embodiment of the present invention;

[0034] Figure 3 is a structural diagram of a system for processing power electronics abnormal data provided by an embodiment of the present invention. Detailed Embodiments

[0035] The following details the embodiments of this application. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar modules or modules with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are only used to explain this application, and cannot be construed as a limitation to the present invention.

[0036] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of this application means the presence of the stated features, integers, steps, operations, modules and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, modules, components and / or their groups. It should be understood that when we say that a module is "connected" or "coupled" to another module, it can be directly connected or coupled to other modules, or there may also be intermediate modules. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0037] To make the objectives, technical solutions and advantages of this application clearer, the following will further describe the embodiments of this application in detail with reference to the accompanying drawings.

[0038] Currently, with the rapid development of power electronic device technology, power electronic equipment is increasingly widely used in all walks of life, but the problem of equipment failures is gradually increasing. Therefore, carrying out research on the early parametric fault diagnosis of power electronic circuits (such as the aging of components such as electrolytic capacitors and inductors) is of great significance for ensuring the reliability of circuit operation and being able to predict the fault risk before the fault occurs to avoid further hazards. The Cerebellar Model Neural Network (CMNN) is a locally approximated neural network. Its structure is simple, superior to the neural network, and has a fast convergence speed. However, the storage space of the CMNN weight coefficients will increase sharply with the increase of the input dimension, and the real-time adjustment ability is not strong. The Fuzzy Cerebellar Model Neural Network (FCMNN) introduces the fuzzy theory. By fuzzifying the input, it makes up for the weakness of the self-adjustment of CMNN, further improving the accuracy and reliability. However, the network structure parameters of FCMNN are manually selected, with a large randomness, which easily leads to low training efficiency, resulting in a decline in network performance and affecting the approximation ability of the network.

[0039] The method for processing abnormal power electronic data provided by this application aims to solve the above technical problems of the prior art.

[0040] The following uses specific embodiments to elaborate in detail on the technical solution of this application and how the technical solution of this application solves the above technical problems. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes may not be elaborated in some embodiments. The embodiments of this application will be described below in conjunction with the accompanying drawings.

[0041] As Figure 1 shown in the flowchart of a method for processing abnormal power electronic data provided by an embodiment of the present invention, it includes:

[0042] Step S101, obtaining the electronic signal of the electronic component to be processed.

[0043] In the embodiment of the present invention, the electronic signal of the electronic component usually includes a current signal and a voltage signal. Therefore, the electronic signal of the electronic component to be processed is usually obtained through the following devices: current sensor, voltage sensor.

[0044] Step S102, converting the electronic signal into a digital signal.

[0045] In the embodiment of the present invention, the electronic signal is not as good as the digital signal in signal processing. Therefore, the electronic signal can be converted into a digital signal through an analog-to-digital conversion circuit.

[0046] Step S103, using a genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronic parameter type fault samples to analyze the digital signal.

[0047] In the embodiment of the present invention, the original GA-FCMNN model cannot analyze the digital signal quickly and accurately. It is necessary to train the original GA-FCMNN model with the power electronic parameter type fault samples obtained in advance. The trained GA-FCMNN can then analyze the digital signal quickly and accurately.

[0048] Step S104, comparing the result of the analysis with a preset normal result. If there is an abnormality, process the electronic component to be processed.

[0049] In the embodiment of the present invention, the result analyzed by the GA-FCMNN model is compared with the preset normal result. If the analyzed result is different from the normal result, in order to ensure safety, the electronic component to be processed is processed.

[0050] The processing of the electronic component to be processed includes: cutting off the power supply of the electronic component to be processed and / or giving an alarm.

[0051] For an abnormal electronic component, the processing system can directly cut off the power supply of the electronic component, which can ensure the safety of the electronic component and personnel in the first place, and can also notify the staff through an alarm.

[0052] In an embodiment of the present invention, the obtained electronic signal of the electronic component is converted into a digital signal, and the GA-FCMNN model trained with power electronic parameter type fault samples is used to analyze the digital signal. The analysis result is compared with the normal result. When the comparison result is abnormal, the electronic component to be processed is processed. Compared with FCMNN, the GA-FCMNN model used introduces a genetic algorithm with better optimization ability, and performs genetic optimization operations when initializing the parameters of the neural network to obtain the best initial weights. The genetic algorithm greatly optimizes the selection of the initial weights of the fuzzy cerebellar model neural network, effectively reduces the blindness and time cost of manually selecting initial parameters, further improves the learning efficiency and intelligent level of the neural network diagnostic classifier, and improves the recognition efficiency of electronic component faults.

[0053] As an alternative embodiment of the present invention, before obtaining the electronic signal of the electronic component to be processed, the processing method further includes:

[0054] Setting an anti-external circuit interference device.

[0055] In an embodiment of the present invention, in order to ensure that the obtained electronic signal of the electronic component to be processed is as accurate as possible, an anti-external circuit interference device can be set to prevent interference from the external circuit. This device is usually: optocoupler isolation.

[0056] As another alternative embodiment of the present invention, before using the genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronic parameter type fault samples to analyze the digital signal, the processing method further includes:

[0057] Training the GA-FCMNN model with power electronic parameter type fault samples.

[0058] The training of the GA-FCMNN model with power electronic parameter type fault samples includes:

[0059] 1. Establish a power electronic parameter type fault sample library and establish a GA-FCMNN model.

[0060] In the embodiments of the present invention, through power electronics simulation research such as Matlab, a simulation circuit is built for power electronic components. Based on the failure principles of each component, parametric faults within a reasonable range are set for the components, and the electronic signal sample data at each measurement point is collected to construct an early training sample library for power electronics faults. The sample library is used for the next step of denoising the sample data, time and frequency domain analysis, extracting the fault features that can reflect the faults, and using them as the input of the GA-FCMNN model to train the GA-FCMNN model. When diagnosing the actual operating circuit, the sample information extracted will also be recorded. When the samples accumulate to a certain number, they can participate in constructing a new sample database to enrich the data in the sample library.

[0061] The main body of the GA-FCMNN model (Genetic Algorithm-Fuzzy Cerebellar Model Neural Network) is the fuzzy cerebellar model neural network. The fuzzy set theory is introduced on the basis of the cerebellar model neural network, enabling the model to not only have the ability to process uncertain information of fuzzy logic but also have good fast adaptive learning ability. Its network structure mainly consists of five parts: a fault feature input layer, an associative memory layer, a receptive field layer, a weight layer, and a fault type output layer. Among them, the fuzzy set theory is introduced in the associative memory layer, and an input state will simultaneously activate multiple fuzzy sets to fuzzify the input, facilitating better processing of uncertain information, and thus forming a fuzzy cerebellar model neural network.

[0062] Since the genetic algorithm has good optimization ability, the genetic algorithm can be used to optimize the selection of the initial weights of the fuzzy cerebellar model neural network. The basic idea is as follows: First, initialize the network weights within a set area, encode the initial weights and construct chromosomes. Set the length of each chromosome according to the number of neural network weights and randomly generate N populations; then, according to the evolutionary principle of "survival of the fittest", take the "minimum network error" as the genetic evolution criterion, and perform genetic operations such as selection, crossover, and mutation in the genetic space to evolve the current chromosome, and thus generate a new generation of populations; repeat the above genetic operations to continuously evolve the initial weights of the network, so as to obtain a set of weights within the set area that can minimize the network error found by the genetic algorithm; then use this set of weights as the initial weights of the neural network to further train the network. Introducing the genetic algorithm into the optimization of the initial weights of the neural network can effectively reduce the blindness and time cost of manually selecting initial parameters and further improve the intelligence of the neural network diagnostic device.

[0063] Preferably, for the parameters of the GA-FCMNN model, they are updated by the Back-Propagation algorithm of gradient descent.

[0064] The input fault feature vector calculates the actual output of the network along the neural network signal propagation direction, then compares the actual output with the expected output, and calculates the objective function. At this time, if the objective function value does not meet the error accuracy, the gradient descent method is used to correct the network weight parameters, and the correction direction is along the reverse direction of the neural network propagation. Then calculate the objective function again, and determine again whether the error accuracy is met, and loop until the error requirement is met; if the objective function value meets the error, the training is completed, and the current weight is recorded as the optimal weight of the classifier.

[0065] Update using the Back-Propagation algorithm, including:

[0066] In GA-FCMNN, the parameters to be updated are , , Three parameters:

[0067] Set the objective function as:

[0068] where K is the Kth training. represents the expected output, represents the output of GA-FCMNN.

[0069] If the objective function value does not meet the error accuracy , then use the gradient descent method to correct and update the weights of GA-FCMNN until the accuracy requirement is met:

[0070] (2)

[0071] where is the parameter matrix, is the learning rate matrix, is expressed as:

[0072]

[0073] The parameter change amount can be obtained through the chain rule as:

[0074]

[0075] 2. Train the GA-FCMNN model with the power electronic parameter type fault samples.

[0076] In the embodiments of the present invention, the training of the GA-FCMNN model is for the early diagnosis and identification of power electronic faults. Electrical signals at each measurement point of the converter are collected through a data acquisition card, fault features are extracted, and these are used as the input to the neural network; the weights of the neural network are encoded and initialized, the population size is initialized, fitness calculation, selection, crossover, and mutation operations are performed, and optimization is carried out until the current optimal initial weights are obtained, which are used as the optimal initial weights of the neural network; GA-FCMNN is trained. Through network calculation, if the error accuracy requirement is not met, the Back-Propagation algorithm with gradient descent is used to update the parameters until the maximum number of training times is reached or the error accuracy requirement is met. At this time, the current optimal weights are obtained, which are used as the optimal weights of the neural network parameters, and are sent to the trained GA-FCMNN for testing to obtain the fault diagnosis situation. If there is a fault, the fault point is located and the fault type is identified.

[0077] As Figure 2 shown in the flowchart of a method for processing power electronic abnormal data provided by another embodiment of the present invention, the process of optimizing the initial value of FCMNN through the GA algorithm includes:

[0078] (1) Train GA-FCMNN with training samples, and the neural network performs forward calculation.

[0079] (2) Calculate the objective function (population fitness).

[0080] Take the objective function of the neural network as the fitness function of the genetic algorithm. The fitness function is used to measure how close each individual in the population is to the optimal solution. The smaller the fitness function value, the higher the individual fitness is considered to be, and the better it is, and it is inherited to the next generation with a higher probability. The calculation formula of its fitness function is

[0081] (7)

[0082] In the formula, is the number of network output nodes; is the expected output of the o-th node in FCMNN; is the predicted output of the o-th node.

[0083] (3) Update the parameters.

[0084] (4) Determine whether the end condition is met. The end condition is based on whether the fitness function value reaches the error accuracy requirement or whether the maximum number of generations of evolution is reached. If so, the optimal initial weights are obtained; if not, genetic operations such as selection, crossover, and mutation are performed to obtain new individuals in the population, and then they are sent back to the GA-FCMNN neural network for forward calculation again until the end condition is met.

[0085] Among them, the genetic operation process:

[0086] Selection: The process of selecting individuals with high fitness in the population and retaining them in the next generation population with a relatively high genetic rate. There are various methods for the selection operation, such as the roulette wheel method and the tournament method. In this study, the roulette wheel method is selected, that is, the selection strategy based on fitness proportion. Each individual 's selection probability is

[0087] (8)

[0088] (9)

[0089] In the formula, is the fitness value of individual . Since the smaller the fitness value, the better, the reciprocal of the fitness value is taken before individual selection; is the number of individuals in the population.

[0090] Crossover: The chromosome crossover combination of individuals in the population, passing on excellent characteristics to the offspring, generating new individuals, further expanding the solution space, and increasing the possibility of finding the optimal solution. Since the individuals use real number coding, the real number crossover method is used for the crossover operation. The crossover operation method for the th chromosome and the th chromosome at the

[0091] (10)

[0092] In the formula, is a random number between [0, 1].

[0093] Mutation: Generate new individuals through mutation to maintain population diversity. Select the th gene of the th individual for mutation. The mutation operation method is as follows:

[0094] (11)

[0095] In the formula, is the upper bound of gene ; is the lower bound of gene ; ; is a random number; is the current iteration number; is the maximum number of generations of evolution; is a random number between [0, 1].

[0096] Construct the GA-FCMNN model, including:

[0097] FCMNN is a neural network based on Gaussian fuzzy function, and its fuzzy rules are as follows:

[0098] If is and is …, and is then

[0099] (12)

[0100] In the formula, is the input dimension, is the input of the layer, the input fuzzy set corresponding to the block of the layer, is the weight corresponding to the output of the

[0101] The fuzzy cerebellar model neural network optimized by genetic algorithm mainly consists of five parts: the fault feature input layer, the associative memory layer, the receptive field layer, the weight layer and the fault type output layer. Specifically as follows:

[0102] (1) The first layer is the fault feature input layer: The fault feature vector of the constructed sample is sent into the GA-FCMNN network as the input of the GA-FCMNN network for forward calculation.

[0103] (2) The second layer is the associative memory layer: Quantize the input fault feature vector, and for each quantization discrete region (usually called an element, or resolution). The Gaussian function is used as the activation function in each block. Therefore, under the action of the activation function, the output relationship of the input corresponding to the layer and the block of the is:

[0104] (13)

[0105] and respectively represent the mean (center) and variance (width) of the Gaussian function.

[0106] (3) The third layer is the receptive field layer: it performs a cumulative multiplication on the activated associative memory area in the second layer to calculate the triggering intensity of the input on the associative unit. is the triggering intensity of the th receptive field of the

[0107]

[0108] The fourth layer is the weight layer: it stores the weights between the receptive field space in the third layer and the th output in the fifth layer in the fourth layer . The weight of each output is expressed as:

[0109]

[0110] (5) The fifth layer is the fault type output layer: it performs an operation on the intensity triggered in the third layer and the weights in the fourth layer, and compresses the output value between (0, 1) through the Sigmoid function (mainly depending on the setting of the output label, and binary coding is adopted in this paper) to obtain the th output :

[0111] (16)

[0112] The difference between the proposed network structure and FCMNN is that in the associative memory layer and the weight layer, for the selection of parameters such as , and , GA-FCMNN uses a genetic algorithm to automatically optimize. New gene individuals are generated through chromosome crossover in the population ( , and combined) to expand the solution space and search for the possibility of the optimal solution. When the fitness value of the new individual meets the error precision requirement, the chromosome corresponding to the network initial weights is obtained to obtain the optimal network initial parameters.

[0113] As Figure 3 shown is the structural diagram of a power electronic abnormal data processing system provided by an embodiment of the present invention, including:

[0114] An electronic signal acquisition module 301, configured to acquire the electronic signal of the electronic component to be processed.

[0115] In the embodiment of the present invention, the electronic signal of the electronic component usually includes a current signal and a voltage signal. Therefore, the electronic signal of the electronic component to be processed is usually acquired through the following devices: a current sensor, a voltage sensor.

[0116] A conversion module 302 for converting the electronic signal into a digital signal.

[0117] In the embodiment of the present invention, the electronic signal is not as good as the digital signal in signal processing. Therefore, the electronic signal can be converted into a digital signal through an analog-to-digital conversion circuit.

[0118] An analysis module 303 for analyzing the digital signal using a genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronic parameter type fault samples.

[0119] In the embodiment of the present invention, the original GA-FCMNN model cannot analyze the digital signal quickly and accurately. It is necessary to train the original GA-FCMNN model with power electronic parameter type fault samples obtained in advance. The trained GA-FCMNN can analyze the digital signal quickly and accurately.

[0120] A processing module 304 for comparing the result of the analysis with a preset normal result. If there is an abnormality, the electronic component to be processed is processed.

[0121] In the embodiment of the present invention, the result analyzed by the GA-FCMNN model is compared with a preset normal result. If the analyzed result is different from the normal result, in order to ensure safety, the electronic component to be processed is processed.

[0122] The processing of the electronic component to be processed includes: cutting off the power supply of the electronic component to be processed and / or giving an alarm.

[0123] For an electronic component with an abnormality, the processing system can directly cut off the power supply of the electronic component, which can ensure the safety of the electronic component and personnel in the first time, and can also notify the staff through an alarm.

[0124] In the embodiment of the present invention, the electronic signal of the obtained electronic component is converted into a digital signal, the digital signal is analyzed using a GA-FCMNN model trained with power electronic parameter type fault samples, the result of the analysis is compared with the normal result, and when the comparison result is abnormal, the electronic component to be processed is processed. Compared with FCMNN, the GA-FCMNN model used introduces a genetic algorithm with better optimization ability. When initializing the parameters of the neural network, a genetic optimization operation is performed to enable it to obtain the best initial weights. The genetic algorithm greatly optimizes the selection of the initial weights of the fuzzy cerebellar model neural network, effectively reducing the blindness and time cost of manually selecting initial parameters, further improving the learning efficiency and intelligent level of the neural network diagnostic classifier, and improving the recognition efficiency of electronic component faults.

[0125] As an alternative embodiment of the present invention, the processing system further includes:

[0126] A setting module for setting an anti-external circuit interference device.

[0127] In the embodiment of the present invention, in order to ensure that the electronic signals of the electronic components to be processed are as accurate as possible, an anti-external circuit interference device can be set to prevent external circuit interference. This device is usually: optocoupler isolation.

[0128] As another alternative embodiment of the present invention, the processing system further includes:

[0129] A training module for training the GA-FCMNN model with power electronic parameter type fault samples.

[0130] The training module includes

[0131] A building unit for building a power electronic parameter type fault sample library and building a GA-FCMNN model.

[0132] In the embodiment of the present invention, through power electronic simulation studies such as Matlab, a simulation circuit is built for power electronic components. Using the failure principles of each component, component parameter type faults within a reasonable range are set, and electronic signal sample data at each measurement point is collected to construct an early training sample library for power electronic faults. The sample library is used for the next step of sample data denoising, time-frequency domain analysis, extracting fault features that can reflect faults, and using them as the input of the GA-FCMNN model to train the GA-FCMNN model. When diagnosing the actual operating circuit, the sample information extracted will also be recorded. When the samples accumulate to a certain number, they can participate in building a new sample database to enrich the data in the sample library.

[0133] The main body of the GA-FCMNN model (Genetic Algorithm-Fuzzy Cerebellar Model Neural Network) is a fuzzy cerebellar model neural network. The fuzzy set theory is introduced on the basis of the cerebellar model neural network, making the model not only have the ability to process uncertain information of fuzzy logic, but also have good fast adaptive learning ability. Its network structure mainly consists of five parts: a fault feature input layer, an associative memory layer, a receptive field layer, a weight layer, and a fault type output layer. Among them, the fuzzy set theory is introduced in the associative memory layer, and an input state will activate multiple fuzzy sets at the same time, making the input fuzzy and facilitating better processing of uncertain information, thus forming a fuzzy cerebellar model neural network.

[0134] Due to the good optimization ability of the genetic algorithm, the genetic algorithm can be used to optimize the selection of the initial weights of the fuzzy cerebellar model neural network. The basic idea is as follows: First, initialize the network weights within a set area, encode the initial weights and construct chromosomes. Set the length of each chromosome according to the number of neural network weights, and randomly generate N populations; then, according to the evolutionary principle of "survival of the fittest", take the "minimum network error" as the genetic evolution criterion, and perform genetic operations such as selection, crossover, and mutation in the genetic space to evolve the current chromosome, and thus generate a new generation of populations; loop the above genetic operations to continuously evolve the initial network weights, so as to obtain a set of weights that can minimize the network error found by the genetic algorithm within the set area; then use this set of weights as the initial weights of the neural network and further train the network. Introducing the genetic algorithm into the optimization of the initial weights of the neural network can effectively reduce the blindness and time cost of manually selecting initial parameters, and further improve the intelligence of the neural network diagnostic tool.

[0135] Preferably, for the parameters of the GA-FCMNN model, they are updated by the Back-Propagation algorithm of gradient descent.

[0136] The input fault feature vector calculates the actual output of the network along the neural network signal propagation direction, then compares the actual output with the expected output, and calculates the objective function. At this time, if the objective function value does not meet the error accuracy, the gradient descent method is used to correct the network weight parameters, and the correction direction is along the reverse direction of the neural network propagation, and the objective function is calculated again, and it is determined again whether the error accuracy is met, and the loop is carried out until the error requirement is met; if the objective function value meets the error, the training is completed, and the current weights are recorded as the optimal weights of the classifier.

[0137] The update by the Back-Propagation algorithm includes:

[0138] In GA-FCMNN, the parameters that need to be updated are , , Three parameters:

[0139] Set the objective function as:

[0140]

[0141] where K is the Kth training. t o (K) represents the expected output, y o (K) represents the output of GA-FCMNN.

[0142] If the objective function value does not meet the error accuracy, the gradient descent method is used to correct and update the GA-FCMNN weights until the accuracy requirement is met:

[0143] (2)

[0144] where α = [m ijk , v ijk , w jko T is a parameter matrix, η = diag[η m , η v , η w is a learning rate matrix, is expressed as:

[0145]

[0146] The change in parameters can be obtained by the chain rule as:

[0147] A training unit for training the GA-FCMNN model with the power electronic parameter-based fault samples.

[0148] In the embodiment of the present invention, the training of the GA-FCMNN model is to learn the characteristics of power electronic faults. The electrical signals at each measurement point of the converter are collected by a data acquisition card, the fault characteristics are extracted, and used as the input of the neural network; the weights of the neural network are encoded, the population size is initialized, the fitness is calculated, selection, crossover, and mutation operations are performed, and optimization is carried out until the current optimal initial weights are obtained and used as the best initial weights of the neural network; the GA-FCMNN is trained. Through network calculation, if the error accuracy requirement is not met, the Back-Propagation algorithm with gradient descent is used for parameter update until the maximum training times are reached or the error accuracy requirement is met, and the current optimal weights are obtained and used as the optimal weights of the neural network parameters, completing the training of the GA-FCMNN fault diagnoser. In online work, the real-time signals at each measurement point sampled are used to extract fault characteristics by the same method as in training, and then sent to the trained GA-FCMNN for testing to obtain the fault diagnosis situation. If there is a fault, the fault point is located and the fault type is identified.

[0149] ​It should be understood that although the steps in the flowchart of the accompanying drawings are shown sequentially according to the indication of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear indication in this text, there is no strict order restriction for the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same moment, but can be executed at different moments, and their execution order is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0150] The above are only some embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for processing abnormal data of power electronics, characterized in that, it includes: Obtain the electronic signal of the electronic component to be processed; Convert the electronic signal into a digital signal; Use the genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronics parameter type fault samples to analyze the digital signal; Compare the result of the analysis with the preset normal result. If there is an abnormality, process the electronic component to be processed; Before using the genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronics parameter type fault samples to analyze the digital signal, the processing method further includes: Establish a power electronics parameter type fault sample library and establish a GA-FCMNN model; Train the GA-FCMNN model with the power electronics parameter type fault samples; For the parameters of the GA-FCMNN model, update them through the Back-Propagation algorithm of gradient descent; The update through the Back-Propagation algorithm of gradient descent includes: In the GA-FCMNN model, the parameters to be updated are m ijk , v ijk , w jko These three parameters; Set the objective function as: where K is the K-th training, t o (K) represents the expected output, y o (K) represents the output of the GA-FCMNN model; If the value of the objective function E(K) does not meet the error accuracy ε, then use the gradient descent method to correct and update the weights of the GA-FCMNN model until the accuracy requirement is met: where α = [m ij，k v i,j w k T is the parameter matrix, η = diag[η m , η v , η w is the learning rate matrix, is expressed as: The change in parameters obtained by the chain rule is:​ The process of optimizing the initial value of FCMNN through the GA algorithm includes: (1) Train the GA-FCMNN model with training samples, and the neural network performs forward calculation; (2) Objective function calculation: Take the objective function of the neural network as the fitness function of the genetic algorithm. The fitness function is used to measure how close each individual in the population is to the optimal solution. The smaller the fitness function value, the higher the individual fitness is considered, the better, and it is inherited to the next generation with a higher probability. Its fitness function calculation formula is: where n o is the number of network output nodes; t o is the expected output of the o-th node in FCMNN; y o is the predicted output of the o-th node; (3) Update the parameters; (4) Determine whether the end condition is met. The end condition is based on whether the fitness function value reaches the error accuracy requirement or whether the maximum number of generations of evolution is reached. If so, obtain the optimal initial weights; if not, select, cross, and mutate to obtain new individuals in the population, and then send them into the GA-FCMNN model to perform forward calculation again until the end condition is met; Among them, the genetic operation process: Selection: Roulette wheel selection is chosen, that is, a selection strategy based on fitness proportion. The selection probability p of each individual i i is f i = 1 / F i where F i is the fitness value of individual i. Since the smaller the fitness value, the better, the reciprocal of the fitness value is taken before individual selection; N is the number of individuals in the population; Crossover: The real number crossover method is adopted. For the k-th chromosome a k and the l-th chromosome a l , the crossover operation method at the j-th position is as follows: where r 1 is a random number between [0, 1]; Mutation: Select the j-th gene a of the i-th individual ij Perform mutation, and the mutation operation method is as follows: where gene a ij has a value range of a max and a min ; f(g) = r 0 (1 - g / G max ) 2 ; r 0 is a random number; g is the current iteration number; G max is the maximum number of generations; r 2 is a random number between [0, 1], and f(g) gradually decreases as the number of generations increases; Construct the GA-FCMNN model, including: FCMNN is a neural network based on Gaussian fuzzy function, and its fuzzy rules are as follows: If I 1 is r 1jk and I 2 is r 2jk …, and I ni is r nijk , then y o = w jko for j = 1, 2, …, n l ; k = 1, 2, …, n b where n i is the input dimension, r ijk is the input fuzzy set corresponding to the i-th input, the j-th layer, and the k-th block, and w jko is the weight corresponding to the j-th layer, the k-th block, and the o-th output; The fuzzy cerebellar model neural network optimized based on the genetic algorithm mainly consists of a fault feature input layer, an associative memory layer, a receptive field layer, a weight layer, and a fault type output layer, specifically as follows: (1) The first layer is the fault feature input layer: the fault feature vector I of the constructed sample i (i = 1, 2,..., n i ) is fed into the GA-FCMNN model as input for forward calculation; (2) The second layer is the associative memory layer: Quantize the input fault feature vector, and for each I of the input vector i Quantize n b discrete regions, and the Gaussian function is used as the activation function in each block. Therefore, under the action of the activation function, the input fuzzy set r corresponding to the i-th input, the j-th layer, and the k-th block ijk is: (3) The third layer is the receptive field layer: It multiplies the activated associative memory areas in the second layer to calculate the triggering intensity of the input on the associative units, and b jk is the triggering intensity of the k-th receptive field in the j-th layer: The fourth layer is the weight layer: store the weights between the receptive field space of the third layer and the o-th output of the fifth layer in the fourth layer W o Among them, the weights of each output are expressed as: (5) The fifth layer is the fault type output layer: It calculates the intensity triggered by the third layer and the weights of the fourth layer, and compresses the output value between (0, 1) through the Sigmoid function to obtain the o-th output y o :

2. The processing method according to claim 1, characterized in that, Before obtaining the electronic signal of the electronic component to be processed, the processing method further includes: Set anti-external circuit interference equipment.

3. The processing method according to claim 2, characterized in that, The processing of the electronic component to be processed includes: cutting off the power supply of the electronic component to be processed and / or giving an alarm.

4. A processing system for abnormal power electronics data, based on the processing method according to any one of claims 1 to 3, characterized in that, the processing system includes: an electronic signal acquisition module, configured to acquire the electronic signal of the electronic component to be processed; a conversion module, configured to convert the electronic signal into a digital signal; an analysis module, configured to analyze the digital signal by using a genetic algorithm optimized fuzzy cerebellar model neural network GA-FCMNN model trained with power electronics parameter type fault samples; a processing module, configured to compare the result of the analysis with a preset normal result, and if there is an abnormality, process the electronic component to be processed; the processing system further includes: a training module, configured to train the GA-FCMNN model with power electronics parameter type fault samples; the training module includes a establishment unit, configured to establish a power electronics parameter type fault sample library and establish a GA-FCMNN model; a training unit, configured to train the GA-FCMNN model with the power electronics parameter type fault samples.

5. The processing system according to claim 4, characterized in that, the processing system further includes: a setting module, configured to set an anti-external circuit interference device.

6. The processing system according to claim 5, characterized in that, processing the electronic component to be processed includes: cutting off the power supply of the electronic component to be processed and / or giving an alarm.

Images