A power battery fault diagnosis method based on an improved long short-term memory neural network

Abstract

The application discloses a power battery fault diagnosis method based on an improved long short-term memory neural network and belongs to the technical field of power battery safety control. The application aims to provide a method for diagnosing faults of a power battery and outputting a health state of the power battery by only requiring a small amount of power battery charging and discharging data, so as to solve many defects of existing methods. On the basis of a long short-term memory neural network, the application uses a genetic algorithm to optimize hyperparameters of a model to obtain an optimal network structure and parameter setting, then uses a lithium battery charging and discharging data set to train, uses a trained GA-LSTM model to diagnose faults, and obtains a health state. Compared with existing power battery fault diagnosis methods, the method has the advantages of higher precision, good stability and strong generalization ability.

Description

Technical Field This invention discloses a power battery fault diagnosis method based on an improved long short-term memory neural network, belonging to the field of power battery safety control technology. Background Technology To ensure good battery performance and the electrical safety of electric vehicles, necessary fault diagnosis is required. Timely provision of battery fault information and health status is essential. The ability to accurately diagnose power battery faults is a crucial indicator for evaluating battery safety management systems. Currently, commonly used battery fault diagnosis methods mainly include traditional modeling and calculation methods and fuzzy neural network diagnostic methods. Traditional modeling and calculation methods rely on the accuracy of the equivalent circuit model, combined with the current operating data of the battery, to calculate and analyze the results. However, this method is limited by calculation speed and model accuracy. While fuzzy neural networks improve calculation speed, the results are presented in fuzzy language, making it impossible to intuitively determine the degree of fault. Summary of the Invention The purpose of this invention is to provide a method for diagnosing faults in power batteries and outputting the health status of power batteries with only a small amount of power battery charge and discharge data, so as to solve many shortcomings of existing methods.

[0004] The power battery fault diagnosis method described in this invention is implemented according to the following steps.

[0005] Step 1: Use a battery testing system to perform a charge and discharge test on the power battery. The sampling time is 30 seconds. The cycle stages are: (1) rest for 120 seconds, (2) charge with a constant 1C current until the voltage reaches the rated charging voltage, (3) charge with the rated voltage until the charging current drops to 20mA, (4) rest for 120 seconds, (5) discharge with a constant 0.5C current until the voltage drops to 75% of the rated voltage, (6) rest for 120 seconds. Record the test time, cycle stage, number of cycles, current, voltage, charging capacity, discharging capacity, and internal resistance data.

[0006] Step 2, extract features from the collected data. Overall voltage level: (1) Calculate the average voltage of each cycle stage, (2) Calculate the percentage of each cycle stage in the whole cycle, (3) Multiply the average voltage of each cycle stage and the percentage and sum them to obtain the overall voltage level. Voltage change rate: (1) Divide the difference between the voltages of adjacent sampling points by the sampling time to obtain the voltage change rate of the sampling point. The formula is: (2) Divide the intervals by the range of the sampling point change rate, count the data points in each interval, and calculate the percentage. (3) Multiply the interval midpoint and the data point distribution percentage to obtain the voltage change rate. Constant current charging time: cumulative time of constant current charging stage. Constant voltage charging time: cumulative time of constant voltage charging stage.

[0007] Step 3: Filter outliers from the collected data. The filtering steps are as follows: (1) Calculate the mean and variance of the data; (2) Define 3 times the variance as the threshold, compare the data with the outlier range, identify data points outside the range and define them as outliers; (3) Delete the outliers and reconstruct the data.

[0008] Step four: To avoid the negative impact of data range on subsequent neural network training and to make the training more effective, the data after outlier filtering is normalized using the min-max normalization method. The formula is as follows:

[0009] Step 5: Optimize the hyperparameters of the Long Short-Term Memory Neural Network using a genetic algorithm. The steps of the genetic algorithm are as follows: (1) Initialize the population and generate a random array, where each number represents an individual; (2) For each individual, evaluate its fitness in the problem to measure its quality; (3) Select excellent individuals for the next generation of the population using the roulette wheel method based on fitness; (4) Randomly select a pair of individuals from the selected individuals for crossover to generate offspring individuals with parental advantages; (5) Perform a certain probability mutation operation on the newly generated individuals to introduce diversity and prevent the algorithm from getting stuck in local optima; (6) Combine the new individuals with the original population until a hyperparameter that significantly improves the prediction accuracy of the Long Short-Term Memory Neural Network is found.

[0010] Step 6: Use Matlab software to build a genetic algorithm-improved long short-term memory neural network, and divide the processed data into training and test sets in an 8:2 ratio.

[0011] Step 7: Using test time, number of cycles, current, and voltage as input features, and charging capacity, discharging capacity, and internal resistance as output features, input the training set into the network and train the neural network using the LM algorithm. After training, input the test set into the network for testing, and calculate the relative error (RE), mean square error (MSE), and root mean square error (RMSE) between the actual values ​​and the network output values. The calculation formulas are as follows:

[0012]

[0013] Step 8: Calculate the average values ​​of the overall voltage level, voltage change rate, constant current charging time, constant voltage charging time, capacity, and internal resistance in the dataset, and use these values ​​as the optimal state of the battery, with a health state of 100%.

[0014] Step 9: Perform a complete single cycle on the battery to be diagnosed, record the diagnostic data, and use a neural network to predict the data, outputting capacity and internal resistance data.

[0015] Step 10: Based on the cycle data of the battery to be diagnosed, calculate the overall voltage level, voltage change rate, constant current charging time, and constant voltage charging time. Then, combine this with the capacity and internal resistance data predicted by the neural network. Calculate the relative proportions of each of the six indicators to the optimal state. Sum all the relative proportions and divide by the optimal comprehensive proportion to obtain the comprehensive index. Finally, convert the comprehensive index into a percentage of healthy state. The calculation formula is as follows: Attached Figure Description Figure 1 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation like Figure 1 As shown, the power battery fault prediction method based on long short-term memory neural network described in this embodiment is implemented according to the following steps.

[0018] Step 1: Use a battery testing system to perform a charge and discharge test on the power battery. The sampling time is 30 seconds. The cycle stages are: (1) rest for 120 seconds, (2) charge with a constant 1C current until the voltage reaches the rated charging voltage, (3) charge with the rated voltage until the charging current drops to 20mA, (4) rest for 120 seconds, (5) discharge with a constant 0.5C current until the voltage drops to 75% of the rated voltage, (6) rest for 120 seconds. Record the test time, cycle stage, number of cycles, current, voltage, charging capacity, discharging capacity, and internal resistance data.

[0019] Step 2, extract features from the collected data. Overall voltage level: (1) Calculate the average voltage of each cycle stage, (2) Calculate the percentage of each cycle stage in the whole cycle, (3) Multiply the average voltage of each cycle stage and the percentage and sum them to obtain the overall voltage level. Voltage change rate: (1) Divide the difference between the voltages of adjacent sampling points by the sampling time to obtain the voltage change rate of the sampling point. The formula is: (2) Divide the intervals by the range of the sampling point change rate, count the data points in each interval, and calculate the percentage. (3) Multiply the interval midpoint and the data point distribution percentage to obtain the voltage change rate. Constant current charging time: cumulative time of constant current charging stage. Constant voltage charging time: cumulative time of constant voltage charging stage.

[0020] Step 3: Filter outliers from the collected data. The filtering steps are as follows: (1) Calculate the mean and variance of the data; (2) Define 3 times the variance as the threshold, compare the data with the outlier range, identify data points outside the range and define them as outliers; (3) Delete the outliers and reconstruct the data.

[0021] Step four: To avoid the negative impact of data range on subsequent neural network training and to make the training more effective, the data after outlier filtering is normalized using the min-max normalization method. The formula is as follows:

[0022] Step 5: Optimize the hyperparameters of the Long Short-Term Memory Neural Network using a genetic algorithm. The steps of the genetic algorithm are as follows: (1) Initialize the population and generate a random array, where each number represents an individual; (2) For each individual, evaluate its fitness in the problem to measure its quality; (3) Select excellent individuals for the next generation of the population using the roulette wheel method based on fitness; (4) Randomly select a pair of individuals from the selected individuals for crossover to generate offspring individuals with parental advantages; (5) Perform a certain probability mutation operation on the newly generated individuals to introduce diversity and prevent the algorithm from getting stuck in local optima; (6) Combine the new individuals with the original population until a hyperparameter that significantly improves the prediction accuracy of the Long Short-Term Memory Neural Network is found.

[0023] Step 6: Use Matlab software to build a genetic algorithm-improved long short-term memory neural network, and divide the processed data into training and test sets in an 8:2 ratio.

[0024] Step 7: Using test time, number of cycles, current, and voltage as input features, and charging capacity, discharging capacity, and internal resistance as output features, input the training set into the network and train the neural network using the LM algorithm. After training, input the test set into the network for testing, and calculate the relative error (RE), mean square error (MSE), and root mean square error (RMSE) between the actual values ​​and the network output values. The calculation formulas are as follows:

[0025]

[0026] Step 8: Calculate the average values ​​of the overall voltage level, voltage change rate, constant current charging time, constant voltage charging time, capacity, and internal resistance in the dataset, and use these values ​​as the optimal state of the battery, with a health state of 100%.

[0027] Step 9: Perform a complete single cycle on the battery to be diagnosed, record the diagnostic data, and use a neural network to predict the data, outputting capacity and internal resistance data.

[0028] Step 10: Based on the cycle data of the battery to be diagnosed, calculate the overall voltage level, voltage change rate, constant current charging time, and constant voltage charging time. Then, combine this with the capacity and internal resistance data predicted by the neural network. Calculate the relative proportions of each of the six indicators to the optimal state. Sum all the relative proportions and divide by the optimal comprehensive proportion to obtain the comprehensive index. Finally, convert the comprehensive index into a percentage of healthy state. The calculation formula is as follows:

Claims

1. A method for fault diagnosis of power batteries based on an improved long short-term memory neural network, characterized in that: The method is implemented according to the following steps: Step 1, the power battery is charged and discharged using a battery testing system. The sampling time is 30 seconds. The cycle stages are divided into: (1) resting for 120 seconds, (2) charging with a constant 1C current until the voltage reaches the rated charging voltage, (3) charging with the rated voltage until the charging current drops to 20mA, (4) resting for 120 seconds, (5) discharging with a constant 0.5C current until the voltage drops to 75% of the rated voltage, (6) resting for 120 seconds. The test time, cycle stage, number of cycles, current, voltage, charging capacity, discharging capacity, and internal resistance data are recorded. Step 2, extract features from the collected data. Overall voltage level: (1) Calculate the average voltage of each cycle stage, (2) Calculate the percentage of each cycle stage in the whole cycle, (3) Multiply the average voltage of each cycle stage and the percentage and sum them to obtain the overall voltage level. Voltage change rate: (1) Divide the difference between the voltages of adjacent sampling points by the sampling time to obtain the voltage change rate of the sampling point. The formula is: (2) Divide the data into intervals based on the range of the sampling point change rate, count the data points in each interval, and calculate the percentage. (3) Multiply the midpoint of the interval and the percentage of the data point distribution to obtain the voltage change rate. Constant current charging time: cumulative time of constant current charging stage. Constant voltage charging time: cumulative time of constant voltage charging stage. Step 3: Screen the collected data for outliers. The screening steps are as follows: (1) Calculate the average value and variance of the data. (2) Define 3 times the variance as the threshold. Compare the data with the outlier range and identify the data points outside the range. Define them as outliers. (3) Delete the outliers and reconstruct the data. Step 4: In order to avoid the adverse effects of the data span on the subsequent neural network training and to make the neural network training more effective, normalize the data after screening outliers using the maximum-minimum normalization method. The formula is: Step 5: Optimize the hyperparameters of the Long Short-Term Memory Neural Network using a genetic algorithm. The steps of the genetic algorithm are: (1) Initialize the population and generate a random array, where each number represents an individual; (2) For each individual, evaluate its fitness in the problem to determine its quality; (3) Select excellent individuals for the next generation of the population using a roulette wheel method based on fitness; (4) Randomly select a pair of individuals from the selected individuals for crossover to generate offspring individuals with parental advantages; (5) Perform a certain probability mutation operation on the newly generated individuals to introduce diversity and prevent the algorithm from getting stuck in local optima; (6) Combine the new individuals with the original population directly... The process begins with finding hyperparameters that significantly improve the prediction accuracy of the Long Short-Term Memory (LSTM) neural network. Step six involves building a genetically-adapted LSM neural network using Matlab software, and dividing the processed data into training and testing sets in an 8:2 ratio. Step seven uses test time, number of cycles, current, and voltage as input features, and charging capacity, discharging capacity, and internal resistance as output features. The training set is input into the network, and the LM algorithm is used to train the neural network. After training, the test set is input into the network for testing, and the relative error (RE), mean square error (MSE), and root mean square error (RMSE) between the actual values ​​and the network output values ​​are calculated using the following formulas: Step 8: Calculate the average values ​​of the overall voltage level, voltage change rate, constant current charging time, constant voltage charging time, capacity, and internal resistance in the dataset. This average value is taken as the battery's optimal state, with a health status of 100%. Step 9: Perform a complete single cycle on the battery to be diagnosed, record the diagnostic data, and use a neural network to predict the data, outputting capacity and internal resistance data. Step 10: Based on the cycle data of the battery to be diagnosed, calculate the overall voltage level, voltage change rate, constant current charging time, and constant voltage charging time. Combine this with the capacity and internal resistance data predicted by the neural network. Calculate the relative proportions of each of the six indicators with the optimal state. Sum all the relative proportions and divide by the optimal comprehensive proportion to obtain a comprehensive index. Finally, convert the comprehensive index into a health status percentage. The calculation formula is:

Images