Thyristor remaining service life prediction method and system based on LSTM network

Abstract

The invention discloses a thyristor remaining service life prediction method and system based on an LSTM network, and the method comprises the steps: carrying out the preprocessing of monitoring data in the working process of a thyristor, constructing a health factor, constructing an LSTM neural network model, and carrying out the optimization of the LSTM neural network model through a gray wolf algorithm. Finally, the optimized intelligent optimization LSTM neural network model is used for carrying out multi-stage life prediction on the transistor, and the limitation of a traditional prediction method in processing long-time sequence data and complex working environments is effectively solved. By combining the grey wolf algorithm and the difference algorithm for intelligent optimization, the technology not only improves the prediction precision, but also enhances the adaptability of the model to the complex and changeable working environment, and ensures the stable operation of the power system.

Description

Technical Field

[0001] The present invention relates to the technical field of remaining service life prediction, and in particular to a method and system for predicting the remaining service life of a thyristor based on an LSTM network. Background Art

[0002] Thyristors are crucial in power systems. As the core component of power control, they can flexibly adjust the voltage and current in the power system to meet various complex power needs, thereby improving the stability of the system. At the same time, thyristors can also optimize the distribution and utilization of electric energy, improve the efficiency of electric energy utilization, and reduce energy waste. Thyristors play a vital role in the stable operation and high-efficiency management of power systems due to their core control functions, improved system stability, improved energy utilization efficiency, wide application fields, high reliability and long life.

[0003] As an important power control element in the power system, the working state of thyristors directly affects the stability and reliability of the power system. By accurately predicting the remaining service life of thyristors, thyristors that are about to fail can be discovered and replaced in time, thereby ensuring the stable operation of the power electronic system and avoiding system failures caused by thyristor failure. Traditional thyristor remaining service life prediction technologies mainly include failure physical model method, empirical formula method and accelerated aging test method.

[0004] Traditional thyristor remaining service life prediction technologies each have their own unique advantages, but they also have certain limitations. The failure physical model method is complex and time-consuming to establish, and it is difficult to adapt to the influence of dynamic conditions and random noise; the empirical formula method is highly dependent on the integrity and representativeness of the test data, and the prediction accuracy and generalization ability are limited; although the accelerated aging test method can shorten the test cycle, there may be differences between the test conditions and the actual working conditions, and it is difficult to determine the acceleration factor. Therefore, in practical applications, it is necessary to comprehensively consider various factors and select appropriate technical methods for comprehensive evaluation and prediction. Summary of the invention

[0005] Based on the problems raised by the above background technology, the purpose of the present invention is to provide a thyristor remaining service life prediction method based on LSTM network. Through thyristor data preprocessing, key parameters characterizing the aging state of thyristors are extracted; then a basic LSTM neural network is constructed, and the gray wolf algorithm and the differential algorithm are mixed to form an intelligent optimization LSTM algorithm; finally, the remaining service life is predicted by the intelligent optimization LSTM regression model, which solves the problem that traditional prediction technology is often difficult to effectively process long time series data, and thus cannot capture the long-term dependencies in the data.

[0006] The present invention is achieved through the following technical solutions:

[0007] The first aspect of the present invention provides a method for predicting the remaining service life of a thyristor based on an LSTM network, comprising the following steps:

[0008] Acquire monitoring data of the thyristor during operation, and extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method;

[0009] Construct an LSTM neural network model for time series prediction, and use the Gray Wolf Algorithm to preliminarily optimize the hyperparameters in the LSTM neural network model;

[0010] The gray wolf algorithm is optimized by using a differential algorithm to obtain an intelligent optimized LSTM neural network model;

[0011] Based on the health factor, the remaining life of the thyristor is divided into different stages using a support vector machine classification algorithm, and the remaining life of the thyristor approaching the end-of-life stage is predicted using the intelligent optimization LSTM neural network model to obtain the predicted remaining life of the thyristor.

[0012] In the above technical solution, the monitoring data of the current, voltage, temperature, etc. of the thyristor during operation are collected, and the health factor that can characterize the aging state of the thyristor is extracted from the external data of the monitorable thyristor using the correlation coefficient analysis method. The health factor includes specific stage data in the on and off states, the off time, etc. According to the results of the correlation analysis, the health factor that can characterize the aging state of the thyristor is defined. These factors can be a single characteristic parameter or a combination of multiple characteristic parameters. If the change trend of the health factor is consistent with the thyristor life decay trend, it means that the health factor is valid.

[0013] A basic LSTM neural network model is constructed for time series prediction. The Gray Wolf Algorithm is used to perform preliminary optimization on the hyperparameters of the LSTM model, such as the number of hidden layer nodes, number of training times, initial learning rate, etc., to improve the prediction accuracy of the model.

[0014] Then, the differential evolution algorithm is used to update the optimal wolf position. Then, according to the wolf pack update mechanism with strong survival mechanism, the position of the entire population is updated, and finally an intelligent optimization LSTM neural network model is formed, which further improves the prediction ability of the LSTM model and can more accurately identify the key features in the time series data, thereby improving the accuracy of the prediction.

[0015] In the early stage of thyristor use, the support vector machine classification algorithm is used to divide the thyristor into stages, and the remaining life of the thyristor is divided into different stages, so as to quickly give the approximate range of the thyristor life. When the thyristor is close to the end of its life, the intelligent optimization LSTM neural network model is used to accurately predict the remaining life of the thyristor. Compared with the existing remaining life prediction, the remaining life prediction method of this method can better adapt to these complex environments by learning complex patterns in historical data, and improve the reliability and stability of the prediction.

[0016] In an optional embodiment, extracting a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method comprises the following steps:

[0017] Randomly select two continuous variables from the monitoring data;

[0018] The Pearson correlation coefficient analysis method was used to conduct correlation analysis on two continuous variables to obtain the Pearson correlation coefficient of the two continuous variables;

[0019] The Spearman correlation coefficient analysis method was used to conduct correlation analysis on the rank size of two continuous variables to obtain the Spearman correlation coefficient of the two continuous variables;

[0020] Kendall correlation coefficient analysis was used to analyze the correlation between two continuous variables and obtain the Kendall correlation coefficient of the two continuous variables;

[0021] The health factors are extracted from the monitoring data by combining the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

[0022] In an optional embodiment, constructing an LSTM neural network model for time series prediction includes the following steps:

[0023] Use deep learning framework to build LSTM neural network model;

[0024] Initialize the LSTM layer of the LSTM neural network model, set the return sequence parameters, and add a Dense layer for output.

[0025] In an optional embodiment, the hyperparameters in the LSTM neural network model are preliminarily optimized using the Gray Wolf Algorithm, including the following steps:

[0026] The hidden layer nodes, training times and initial learning rate in the LSTM neural network model are used as hyperparameters to be optimized;

[0027] Determine the initial parameters of the gray wolf algorithm, the initial parameters including the search space, the maximum number of iterations, the dimension and the size of the gray wolf population, wherein the search space is the value range of all hyperparameters in the LSTM neural network model;

[0028] Initialize the gray wolf position vector of the gray wolf population, iteratively calculate the optimal fitness value of the gray wolf position vector, and assign the gray wolf position vector corresponding to the optimal fitness value after the iteration to the hyperparameter to be optimized.

[0029] In an optional embodiment, the gray wolf algorithm is optimized using a differential algorithm, comprising the following steps:

[0030] Initialize the population in the grey wolf algorithm;

[0031] Randomly select three different original individuals from the population as individuals to be mutated, perform mutation operations on the individuals to be mutated by using a differential algorithm, and generate mutant individuals;

[0032] Randomly select an original individual from the population as the individual to be crossed, and perform a crossover operation on the variant individual and the individual to be crossed to generate a crossover individual;

[0033] The fitness values ​​of the crossover individual and the original individual are compared according to the greedy criterion, and the individual with the smaller fitness value between the crossover individual and the original individual is used as the original individual of the next generation population.

[0034] In an optional embodiment, three different original individuals are randomly selected from the population as individuals to be mutated, and a mutation operation is performed on the individuals to be mutated by a differential algorithm, including the following steps:

[0035] Select the ath original individual x from the population aj (t), the bth original individual x bj (t) and the cth original individual x cj (t) as the individual to be mutated;

[0036] Performing a mutation operation on the individual to be mutated using a mutation operation formula to generate a mutated individual;

[0037] The mutation operation formula is as follows:

[0038] ν ij (t+1)=x aj (t)+F·(x bj (t)-x cj (t))a≠b≠c

[0039] In the above formula, F represents the mutation operator, ν ij(t+1) represents the i-th mutant individual of the j-th characteristic parameter in the t+1-th iteration process.

[0040] In an optional embodiment, an original individual is randomly selected from the population as an individual to be crossed, and a crossover operation is performed between the variant individual and the individual to be crossed, comprising the following steps:

[0041] Select the i-th original individual x from the population ij (t) as the individual to be crossed;

[0042] A crossover operator is used to perform a crossover operation on the individual to be crossed and the mutated individual. The process of the crossover operation is as follows:

[0043]

[0044] In the above formula, rand(0,1) means generating a random value between 0 and 1, C R represents a crossover operator with a specific value, ranging from 0 to 1, j rand Represents a randomly selected dimension.

[0045] In an optional embodiment, the remaining life of the thyristor is divided into different stages based on the health factor using a support vector machine classification algorithm, comprising the following steps:

[0046] Acquiring monitoring data of the thyristor during operation, and generating a thyristor key parameter curve using the monitoring data;

[0047] Extracting signal energy, signal fluctuation index and signal skewness from the thyristor key parameter curve based on the health factor;

[0048] The support vector machine classification algorithm is used to divide the remaining life of the thyristor into Class A life stage, Class B life stage, Class C life stage and Class D life stage according to the signal energy, the signal fluctuation index and the signal skewness; wherein the Class D life stage is a stage close to the end of life.

[0049] The second aspect of the present invention provides a thyristor remaining service life prediction system based on an LSTM network, comprising:

[0050] A correlation coefficient analysis module, used to obtain monitoring data of the thyristor during operation, and to extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method;

[0051] An initial optimization module, used to construct an LSTM neural network model for time series prediction, and to perform preliminary optimization of hyperparameters in the LSTM neural network model using the Gray Wolf Algorithm;

[0052] An intelligent optimization module, used to optimize the gray wolf algorithm using a differential algorithm to obtain an intelligent optimized LSTM neural network model;

[0053] The prediction module is used to divide the remaining life of the thyristor into different stages based on the health factor using a support vector machine classification algorithm, and use the intelligent optimization LSTM neural network model to predict the remaining life of the thyristor approaching the end-of-life stage to obtain the predicted remaining life of the thyristor.

[0054] In an optional embodiment, the correlation coefficient analysis module includes:

[0055] A value taking unit, used for taking any two continuous variables from the monitoring data;

[0056] A Pearson correlation coefficient analysis unit, used to perform correlation analysis on two continuous variables using the Pearson correlation coefficient analysis method to obtain the Pearson correlation coefficient of the two continuous variables;

[0057] The Spearman correlation coefficient analysis unit is used to perform correlation analysis on the ranks of two continuous variables using the Spearman correlation coefficient analysis method to obtain the Spearman correlation coefficient of the two continuous variables;

[0058] Kendall correlation coefficient analysis unit, used to perform correlation analysis on the correlation between two continuous variables using Kendall correlation coefficient analysis method, and obtain the Kendall correlation coefficient of the two continuous variables;

[0059] The extraction unit is used to extract health factors from the monitoring data by comprehensively analyzing the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0061] 1. LSTM network can learn and memorize long-term dependencies in time series data through its unique gating mechanism. This is particularly important for the prediction of the remaining life of devices such as thyristors;

[0062] 2. By introducing improved methods such as the Grey Wolf Algorithm and the Differential Algorithm, the intelligently optimized LSTM network can more accurately identify key features in time series data, thereby improving the accuracy of predictions;

[0063] 3. The intelligently optimized LSTM neural network model can better adapt to these complex environments and improve the reliability and stability of predictions by learning complex patterns in historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0065] Figure 1 A schematic diagram of a flow chart of a method for predicting the remaining useful life of a thyristor based on an LSTM network provided in Embodiment 1 of the present invention;

[0066] Figure 2 A schematic diagram of the structure of an LSTM network provided in Example 1 of the present invention;

[0067] Figure 3 Prediction block diagram of the gray wolf algorithm provided in Example 1 of the present invention

[0068] Figure 4 A schematic diagram of the flow of the LSTM network algorithm provided in Example 1 of the present invention;

[0069] Figure 5 This is a graph of prediction results of the LSTM network algorithm provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0070] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0071] Example 1

[0072] Figure 1 A schematic diagram of a flow chart of a method for predicting the remaining useful life of a thyristor based on an LSTM network provided in Embodiment 1 of the present invention is shown in FIG. Figure 1 As shown, the thyristor remaining service life prediction method based on LSTM network includes the following steps:

[0073] Acquire monitoring data of the thyristor during operation, and extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method;

[0074] Construct an LSTM neural network model for time series prediction, and use the Gray Wolf Algorithm to preliminarily optimize the hyperparameters in the LSTM neural network model;

[0075] The gray wolf algorithm is optimized by using a differential algorithm to obtain an intelligent optimized LSTM neural network model;

[0076] Based on the health factor, the remaining life of the thyristor is divided into different stages using a support vector machine classification algorithm, and the remaining life of the thyristor approaching the end-of-life stage is predicted using the intelligent optimization LSTM neural network model to obtain the predicted remaining life of the thyristor.

[0077] It should be noted that the monitoring data of the current, voltage, temperature, etc. of the thyristor during operation are collected, and the health factor that can characterize the aging state of the thyristor is extracted from the external data of the monitorable thyristor using the correlation coefficient analysis method. The health factor includes specific stage data in the on and off states, the off time, etc. According to the results of the correlation analysis, the health factor that can characterize the aging state of the thyristor is defined. These factors can be a single characteristic parameter or a combination of multiple characteristic parameters. If the change trend of the health factor is consistent with the thyristor life decay trend, it means that the health factor is valid.

[0078] A basic LSTM neural network model is constructed for time series prediction. The Gray Wolf Algorithm is used to perform preliminary optimization on the hyperparameters of the LSTM model, such as the number of hidden layer nodes, number of training times, initial learning rate, etc., to improve the prediction accuracy of the model.

[0079] Then, the differential evolution algorithm is used to update the optimal wolf position. Then, according to the wolf pack update mechanism with strong survival mechanism, the position of the entire population is updated, and finally an intelligent optimization LSTM neural network model is formed, which further improves the prediction ability of the LSTM model and can more accurately identify the key features in the time series data, thereby improving the accuracy of the prediction.

[0080] In the early stage of thyristor use, the support vector machine classification algorithm is used to divide the thyristor into stages, and the remaining life of the thyristor is divided into different stages, so as to quickly give the approximate range of the thyristor life. When the thyristor is close to the end of its life, the intelligent optimization LSTM neural network model is used to accurately predict the remaining life of the thyristor. Compared with the existing remaining life prediction, the remaining life prediction method of this method can better adapt to these complex environments by learning complex patterns in historical data, and improve the reliability and stability of the prediction.

[0081] In an optional embodiment, extracting a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method comprises the following steps:

[0082] Randomly select two continuous variables from the monitoring data;

[0083] The Pearson correlation coefficient analysis method was used to conduct correlation analysis on two continuous variables to obtain the Pearson correlation coefficient of the two continuous variables;

[0084] The Spearman correlation coefficient analysis method was used to conduct correlation analysis on the rank size of two continuous variables to obtain the Spearman correlation coefficient of the two continuous variables;

[0085] Kendall correlation coefficient analysis was used to analyze the correlation between two continuous variables and obtain the Kendall correlation coefficient of the two continuous variables;

[0086] The health factors are extracted from the monitoring data by combining the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

[0087] It should be noted that the necessary monitoring data, including current, voltage, temperature, time and other parameters, are collected from the switching experiment or actual operation data of the thyristor. These data are usually obtained through special thyristor parameter testers or trigger characteristic testers and other equipment. Check and process outliers, missing values ​​or noise points in the data. This can be achieved by filtering, interpolation or deleting abnormal data points, and normalizing the data to ensure that data of different dimensions can be compared and analyzed on the same scale. Divide the collected data into training set and test set. The training set is used to train the prediction model, and the test set is used to evaluate the prediction performance of the model.

[0088] The correlation coefficient analysis methods such as Pearson, Spearman and Kendall are used to conduct correlation analysis between the extracted features and the thyristor key parameter curve. The purpose is to find the characteristic parameters that are highly correlated with the thyristor key parameter curve, which will serve as the basis for constructing the health factor.

[0089] Among them, the Pearson correlation coefficient can be used to evaluate the linear relationship between two continuous variables. The Pearson correlation coefficient analysis method is used to perform correlation analysis on two continuous variables to obtain the Pearson correlation coefficient of the two continuous variables. Specifically, for the variable x = (x1, x2, ... x n ) and variable y=(y1,y2,…y n ), the Pearson correlation coefficient is calculated as follows:

[0090]

[0091] In the above formula, ρ is the Pearson correlation coefficient and n is the number of variables.

[0092] The Spearman correlation coefficient uses the rank order of two continuous variables for linear correlation analysis. The Spearman correlation coefficient is calculated as follows:

[0093]

[0094] In the above formula, x i and i is the i-th element of two sequences of the same length; rg(x i ) and rg(y i ) is x i and i The ranking of x and y after sorting, where the correlation coefficient is [-1,1], s is the Spearman correlation coefficient, d i Poor ranking.

[0095] Kendall correlation coefficient is used to measure whether two variables are associated. The Kendall correlation coefficient is calculated as follows:

[0096]

[0097] In the above formula, N C The number of pairs of elements representing the consistency between the two sequences; N D It represents the number of inconsistent element pairs in two sequences, and k represents the Kendall correlation coefficient.

[0098] In an optional embodiment, constructing an LSTM neural network model for time series prediction includes the following steps:

[0099] Use deep learning framework to build LSTM neural network model;

[0100] Initialize the LSTM layer of the LSTM neural network model, set the return sequence parameters, and add a Dense layer for output.

[0101] In an optional embodiment, the hyperparameters in the LSTM neural network model are preliminarily optimized using the Gray Wolf Algorithm, including the following steps:

[0102] The hidden layer nodes, training times and initial learning rate in the LSTM neural network model are used as hyperparameters to be optimized;

[0103] Determine the initial parameters of the gray wolf algorithm, the initial parameters including the search space, the maximum number of iterations, the dimension and the size of the gray wolf population, wherein the search space is the value range of all hyperparameters in the LSTM neural network model;

[0104] Initialize the gray wolf position vector of the gray wolf population, iteratively calculate the optimal fitness value of the gray wolf position vector, and assign the gray wolf position vector corresponding to the optimal fitness value after the iteration to the hyperparameter to be optimized.

[0105] It should be noted that if Figure 3As shown in the figure, we first use the deep learning framework to build an LSTM neural network model, and determine the number of LSTM layers of the LSTM neural network model, the number of neurons in each layer, and whether other layers need to be added to output the results. The constructed LSTM structure diagram is shown in the figure below. Figure 2 shown.

[0106] The hyperparameters that need to be optimized for the LSTM algorithm are determined as the hyperparameters to be optimized. The hyperparameters to be optimized are the hidden layer nodes, the number of training times, and the initial learning rate.

[0107] Then determine the four initial parameters of the Gray Wolf Algorithm, search space, maximum number of iterations t max , dimension D, and the size of the gray wolf population N. The search space is the value range of all hyperparameters.

[0108] Initialize the position vector x of each gray wolf in the population i , let the current population be x={x1,x2,…,x N} T , x ij represents the jth dimension of the i-th individual in the population, j = 1, 2, ..., D. Assign the position vector of each gray wolf in the population to the three parameters that need to be optimized by LSTM, and calculate the fitness value of the position vector. The calculation expression of the fitness value is as follows:

[0109]

[0110] Where: f′ represents the fitness value, y′ i Represents the predicted value of the LSTM model; y i Indicates the expected output value.

[0111] Filter out the gray wolf position vectors corresponding to the three smallest fitness values ​​in the population and assign them to x a (t), x β (t), x δ (t), so that it always keeps the fitness minimum. Determine whether the maximum number of iterations has been reached. If so, end the update. After the iteration, assign the optimal x position vector to the hidden layer node, training number, and initial learning rate of the LSTM neural network, and perform model training and prediction according to these parameters.

[0112] In an optional embodiment, the gray wolf algorithm is optimized using a differential algorithm, comprising the following steps:

[0113] Initialize the population in the grey wolf algorithm;

[0114] Randomly select three different original individuals from the population as individuals to be mutated, perform mutation operations on the individuals to be mutated by using a differential algorithm, and generate mutant individuals;

[0115] Randomly select an original individual from the population as the individual to be crossed, and perform a crossover operation on the variant individual and the individual to be crossed to generate a crossover individual;

[0116] The fitness values ​​of the crossover individual and the original individual are compared according to the greedy criterion, and the individual with the smaller fitness value between the crossover individual and the original individual is used as the original individual of the next generation population.

[0117] It should be noted that the process of building an intelligent optimization LSTM neural network model is as follows Figure 4 As shown, the population is initialized, the population size is N, the spatial dimension is D, the evolutionary generation is t, and the current population is x(t) = {x1(t), x2(t), …, x N (t)}. Then the i-th individual in the population can be expressed as x i (t) = {x i1 (t),x i2 (t),…,x iD (t)}, assuming that the boundary condition of the population is Then the jth parameter of the initialized i-th individual can be expressed as formula (5), where j = {1, 2, ..., D} represents the parameter index.

[0118]

[0119] In the above formula: rand[0,1] represents a random number between 0 and 1.

[0120] The mutation operation can generate a new individual through the differential strategy from three different individuals in the population. In order to increase the diversity of the population, the new variable after mutation is used to perform a crossover operation with the original variable to obtain a better individual. The differential evolution algorithm uses the greedy criterion to generate the new crossover individual u i (t+1) and the original individual x i (t) is compared, and the individuals with smaller fitness values ​​are selected as members of the next generation population. The selection process is as follows:

[0121]

[0122] Where: f is the fitness function.

[0123] The basic idea of ​​the intelligent optimization LSTM neural network model is to perform encirclement, hunting and attack behaviors according to the gray wolf algorithm. Then, the differential evolution algorithm is used to update the optimal wolf position. Then, the position of the entire population is updated according to the wolf pack update mechanism of the strong survival mechanism.

[0124] In an optional embodiment, three different original individuals are randomly selected from the population as individuals to be mutated, and a mutation operation is performed on the individuals to be mutated by a differential algorithm, including the following steps:

[0125] Select the ath original individual x from the population aj (t), the bth original individual x bj (t) and the cth original individual x cj (t) as the individual to be mutated;

[0126] Performing a mutation operation on the individual to be mutated using a mutation operation formula to generate a mutated individual;

[0127] The mutation operation formula is as follows:

[0128] ν ij (t+1)=x aj (t)+F·(x bj (t)-x cj (t)) a≠b≠c

[0129] In the above formula, F represents the mutation operator, ν ij (t+1) represents the i-th mutant individual of the j-th characteristic parameter in the t+1-th iteration process.

[0130] It should be noted that x aj (t), x bj (t), x cj (t) are three randomly selected individuals that are different from each other; F is the mutation operator that controls the amplification of the deviation variable, and its value is generally [0,2].

[0131] In an optional embodiment, an original individual is randomly selected from the population as an individual to be crossed, and a crossover operation is performed between the variant individual and the individual to be crossed, comprising the following steps:

[0132] Select the i-th original individual x from the population ij (t) as the individual to be crossed;

[0133] A crossover operator is used to perform a crossover operation on the individual to be crossed and the mutated individual. The process of the crossover operation is as follows:

[0134]

[0135] In the above formula, rand(0,1) means generating a random value between 0 and 1, C R represents a crossover operator with a specific value, ranging from 0 to 1, j rand Represents a randomly selected dimension.

[0136] It should be noted that C R represents a crossover operator with a specific value, ranging from 0 to 1, C R The larger the value, the value v ij The greater the probability of (t+1), the greater the possibility of crossover. R The smaller the value, the more conducive it is to maintaining population diversity and global search.

[0137] In an optional embodiment, the remaining life of the thyristor is divided into different stages based on the health factor using a support vector machine classification algorithm, comprising the following steps:

[0138] Acquiring monitoring data of the thyristor during operation, and generating a thyristor key parameter curve using the monitoring data;

[0139] Extracting signal energy, signal fluctuation index and signal skewness from the thyristor key parameter curve based on the health factor;

[0140] The support vector machine classification algorithm is used to divide the remaining life of the thyristor into Class A life stage, Class B life stage, Class C life stage and Class D life stage according to the signal energy, the signal fluctuation index and the signal skewness; wherein the Class D life stage is a stage close to the end of life.

[0141] It should be noted that as the thyristor continues to age, its current, voltage, temperature and other data can express a nonlinear mapping relationship with the life span in a sense, and important characteristic parameters are extracted from the thyristor key parameter curve. This embodiment extracts three characteristic parameters from the thyristor key parameter curve based on the health factor that can characterize the aging state of the thyristor: signal energy, signal fluctuation index and signal skewness.

[0142] Among them, the signal energy E represents the change of signal strength over time, and the calculation formula is as follows:

[0143]

[0144] In the above formula, x(t) represents the parameter signal.

[0145] The signal fluctuation index FI represents the deviation of the signal from the average value, and the calculation formula is as follows:

[0146]

[0147] In the above formula, x i represents parameter signal; μ is the respective average value; ω represents sampling frequency.

[0148] Signal skewness SI represents the degree to which the signal tends to the signal mean, and the calculation formula is as follows:

[0149]

[0150] In the above formula, μ represents the signal average value; σ represents the signal standard deviation.

[0151] Using the support vector machine (SVM) classification algorithm, the entire life cycle of the thyristor is divided into four stages (A, B, C, and D) according to the extracted characteristic parameters. Each stage represents a different remaining service life percentage range. Class A represents 100%-75%, Class B represents 75%-50%, Class C represents 50%-25%, and Class D represents 25%-0%. The remaining service life RUL is defined as:

[0152]

[0153] In the above formula, N EOL N is the number of cycles of the thyristor from the beginning of use to the life threshold; i The number of cycles for the current state.

[0154] The data of Class D thyristor samples is used to train the intelligent optimization LSTM algorithm regression model, so that it can accurately predict the remaining service life of the battery to improve the prediction accuracy and convergence speed. The prediction results of the intelligent optimization LSTM algorithm are compared with the experimental results. Figure 5 shown.

[0155] Example 2

[0156] Embodiment 2 of the present invention provides a thyristor remaining service life prediction system based on an LSTM network, comprising:

[0157] A correlation coefficient analysis module, used to obtain monitoring data of the thyristor during operation, and to extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method;

[0158] An initial optimization module, used to construct an LSTM neural network model for time series prediction, and to perform preliminary optimization of hyperparameters in the LSTM neural network model using the Gray Wolf Algorithm;

[0159] An intelligent optimization module, used to optimize the gray wolf algorithm using a differential algorithm to obtain an intelligent optimized LSTM neural network model;

[0160] The prediction module is used to divide the remaining life of the thyristor into different stages based on the health factor using a support vector machine classification algorithm, and use the intelligent optimization LSTM neural network model to predict the remaining life of the thyristor approaching the end-of-life stage to obtain the predicted remaining life of the thyristor.

[0161] In an optional embodiment, the correlation coefficient analysis module includes:

[0162] A value taking unit, used for taking any two continuous variables from the monitoring data;

[0163] A Pearson correlation coefficient analysis unit, used to perform correlation analysis on two continuous variables using the Pearson correlation coefficient analysis method to obtain the Pearson correlation coefficient of the two continuous variables;

[0164] The Spearman correlation coefficient analysis unit is used to perform correlation analysis on the ranks of two continuous variables using the Spearman correlation coefficient analysis method to obtain the Spearman correlation coefficient of the two continuous variables;

[0165] Kendall correlation coefficient analysis unit, used to perform correlation analysis on the correlation between two continuous variables using Kendall correlation coefficient analysis method, and obtain the Kendall correlation coefficient of the two continuous variables;

[0166] The extraction unit is used to extract health factors from the monitoring data by comprehensively analyzing the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

[0167] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A thyristor remaining service life prediction method based on LSTM network, characterized in that: The steps include: Acquire monitoring data of the thyristor during operation, and extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method; Construct an LSTM neural network model for time series prediction, and use the Gray Wolf Algorithm to preliminarily optimize the hyperparameters in the LSTM neural network model; The gray wolf algorithm is optimized by using a differential algorithm to obtain an intelligent optimized LSTM neural network model; Based on the health factor, the remaining life of the thyristor is divided into different stages using a support vector machine classification algorithm, and the remaining life of the thyristor approaching the end-of-life stage is predicted using the intelligent optimization LSTM neural network model to obtain the predicted remaining life of the thyristor.

2. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 1, characterized in that: Extracting a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method includes the following steps: Randomly select two continuous variables from the monitoring data; The Pearson correlation coefficient analysis method was used to conduct correlation analysis on two continuous variables to obtain the Pearson correlation coefficient of the two continuous variables; The Spearman correlation coefficient analysis method was used to conduct correlation analysis on the rank size of two continuous variables to obtain the Spearman correlation coefficient of the two continuous variables; Kendall correlation coefficient analysis was used to analyze the correlation between two continuous variables and obtain the Kendall correlation coefficient of the two continuous variables; The health factors are extracted from the monitoring data by combining the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

3. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 1, characterized in that: Building an LSTM neural network model for time series prediction includes the following steps: Use deep learning framework to build LSTM neural network model; Initialize the LSTM layer of the LSTM neural network model, set the return sequence parameters, and add a Dense layer for output.

4. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 3, characterized in that: The Grey Wolf Algorithm is used to preliminarily optimize the hyperparameters in the LSTM neural network model, including the following steps: The hidden layer nodes, training times and initial learning rate in the LSTM neural network model are used as hyperparameters to be optimized; Determine the initial parameters of the gray wolf algorithm, the initial parameters including the search space, the maximum number of iterations, the dimension and the size of the gray wolf population, wherein the search space is the value range of all hyperparameters in the LSTM neural network model; Initialize the gray wolf position vector of the gray wolf population, iteratively calculate the optimal fitness value of the gray wolf position vector, and assign the gray wolf position vector corresponding to the optimal fitness value after the iteration to the hyperparameter to be optimized.

5. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 1, characterized in that: The gray wolf algorithm is optimized by using a differential algorithm, including the following steps: Initialize the population in the grey wolf algorithm; Randomly select three different original individuals from the population as individuals to be mutated, perform mutation operations on the individuals to be mutated by using a differential algorithm, and generate mutant individuals; Randomly select an original individual from the population as the individual to be crossed, and perform a crossover operation on the variant individual and the individual to be crossed to generate a crossover individual; The fitness values ​​of the crossover individual and the original individual are compared according to the greedy criterion, and the individual with the smaller fitness value between the crossover individual and the original individual is used as the original individual of the next generation population.

6. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 5, characterized in that: Randomly select three different original individuals from the population as individuals to be mutated, and perform mutation operations on the individuals to be mutated by using a differential algorithm, including the following steps: Select the ath original individual x from the population aj (t), the bth original individual x bj (t) and the cth original individual x cj (t) as the individual to be mutated; Performing a mutation operation on the individual to be mutated using a mutation operation formula to generate a mutated individual; The mutation operation formula is as follows: ν ij (t+1)=x aj (t)+F·(x bj (t)-x cj (t))a≠b≠c In the above formula, F represents the mutation operator, ν ij (t+1) represents the i-th mutant individual of the j-th characteristic parameter in the t+1-th iteration process.

7. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 6, characterized in that: An original individual is randomly selected from the population as an individual to be crossed, and a crossover operation is performed between the variant individual and the individual to be crossed, comprising the following steps: Select the i-th original individual x from the population ij (t) as the individual to be crossed; A crossover operator is used to perform a crossover operation on the individual to be crossed and the mutated individual. The process of the crossover operation is as follows: In the above formula, rand(0,1) means generating a random value between 0 and 1, C R represents a crossover operator with a specific value, ranging from 0 to 1, j rand Represents a randomly selected dimension.

8. The method for predicting the remaining useful life of a thyristor based on an LSTM network according to claim 1, characterized in that: Based on the health factor, the remaining life of the thyristor is divided into different stages using a support vector machine classification algorithm, including the following steps: Acquiring monitoring data of the thyristor during operation, and generating a thyristor key parameter curve using the monitoring data; Extracting signal energy, signal fluctuation index and signal skewness from the thyristor key parameter curve based on the health factor; The support vector machine classification algorithm is used to divide the remaining life of the thyristor into Class A life stage, Class B life stage, Class C life stage and Class D life stage according to the signal energy, the signal fluctuation index and the signal skewness; wherein the Class D life stage is a stage close to the end of life.

9. The thyristor remaining service life prediction system based on LSTM network is characterized by: include: A correlation coefficient analysis module, used to obtain monitoring data of the thyristor during operation, and to extract a health factor for characterizing the aging state of the thyristor from the monitoring data using a correlation coefficient analysis method; An initial optimization module, used to construct an LSTM neural network model for time series prediction, and to perform preliminary optimization of hyperparameters in the LSTM neural network model using the Gray Wolf Algorithm; An intelligent optimization module, used to optimize the gray wolf algorithm using a differential algorithm to obtain an intelligent optimized LSTM neural network model; The prediction module is used to divide the remaining life of the thyristor into different stages based on the health factor using a support vector machine classification algorithm, and use the intelligent optimization LSTM neural network model to predict the remaining life of the thyristor approaching the end-of-life stage to obtain the predicted remaining life of the thyristor.

10. The thyristor remaining service life prediction system based on LSTM network according to claim 9, characterized in that: The correlation coefficient analysis module comprises: A value taking unit, used for taking any two continuous variables from the monitoring data; A Pearson correlation coefficient analysis unit, used to perform correlation analysis on two continuous variables using the Pearson correlation coefficient analysis method to obtain the Pearson correlation coefficient of the two continuous variables; The Spearman correlation coefficient analysis unit is used to perform correlation analysis on the ranks of two continuous variables using the Spearman correlation coefficient analysis method to obtain the Spearman correlation coefficient of the two continuous variables; Kendall correlation coefficient analysis unit, used to perform correlation analysis on the correlation between two continuous variables using Kendall correlation coefficient analysis method, and obtain the Kendall correlation coefficient of the two continuous variables; The extraction unit is used to extract health factors from the monitoring data by comprehensively analyzing the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient.

Images