A battery thermal runaway risk prediction method and system based on recurrent neural network
Through the battery thermal runaway risk prediction method based on recurrent neural network, deep learning and multi-dimensional feature data are used to solve the problem of inaccurate prediction of battery thermal runaway risk in the prior art, and accurate and timely prediction of battery thermal runaway risk is achieved.
Patent Information
- Application Number
- CN202111661180.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-30
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2041-12-30
AI Technical Summary
The existing battery thermal runaway risk prediction methods are not accurate and timely enough to effectively predict the battery thermal runaway risk.
Using a battery thermal runaway risk prediction method based on recurrent neural network, select feature data from historical data to clean and generate sample data, build a recurrent neural network model, and use the data of normal vehicles and problem vehicles for training and verification, and finally according to Real-time data calculates the deviation degree to determine whether the battery has a risk of thermal runaway.
Accurate and timely prediction of the risk of thermal runaway from the battery is achieved, and the cyclic neural network model is constructed through deep learning methods using characteristic data from multiple dimensions (such as SOC, current, single voltage and temperature), which improves the accuracy and reliability of predictions. .
Smart Images

Figure CN114355199B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of thermal runaway risk monitoring of electric vehicle power batteries, and specifically relates to a battery thermal runaway risk prediction technology based on a recurrent neural network. Background Art
[0002] As the core component of electric vehicles, the performance of power batteries directly determines most of the key performance of electric vehicles. With the gradual popularization of electric vehicles, various problems caused by power batteries have become more and more prominent and have received widespread attention. In order to reduce the failure level of power batteries, in addition to selecting better raw materials, adopting more advanced production processes and stricter quality control in the production and manufacturing process of batteries, it is also necessary to be able to detect the risk of thermal runaway of batteries in advance during the use of batteries to prevent them from happening.
[0003] At present, battery fault warning is mainly divided into two methods: local BMS warning and big data online warning. With the widespread application of big data technology, online warning methods based on big data are receiving more and more attention. The Chinese patent document with publication number CN109978229A discloses a technology called "A method for predicting thermal runaway of multi-point temperature and connection point temperature of all cells in a power battery pack". It mainly collects real-time data of the vehicle, establishes models and training verification algorithms through big data machine learning, establishes XGBoost models, Weibull distribution and Bayesian network models, etc., and predicts thermal runaway of electric vehicle batteries based on the temperature characteristics obtained by the model. This technology mainly predicts battery thermal runaway based on temperature, and the prediction parameters are single.
[0004] The Chinese patent document with publication number CN110161414B discloses a technology named "A method and system for online prediction of thermal runaway of power batteries". It mainly calculates the voltage deviation matrix at each moment based on the voltage value of each battery cell in the power battery, and then inputs the voltage deviation growth rate of each cell corresponding to the current vehicle mileage, the current temperature average of the temperature probe and the voltage deviation growth rate matrix at the current moment T into the thermal runaway cell prediction model to obtain the prediction result of thermal runaway of the power battery. This technology mainly focuses on how to construct the voltage deviation growth rate matrix, and does not provide a specific method for constructing the model. Summary of the invention
[0005] The purpose of the present invention is to provide a battery thermal runaway risk prediction method and system based on a recurrent neural network, and to solve the technical problem that there is no good method for predicting the battery thermal runaway risk in the related art, resulting in the inability to accurately and timely predict the battery thermal runaway risk.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting the risk of thermal runaway of a battery based on a recurrent neural network, comprising the following steps:
[0007] S01: Selecting feature data from historical data collected in advance for training a recurrent neural network model, and performing data cleaning on the feature data;
[0008] S02: generating sample data of a fixed time length based on the characteristic data after cleaning, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle;
[0009] S03: construct the recurrent neural network model;
[0010] S04: defining the output of the recurrent neural network model;
[0011] S05: The recurrent neural network model obtains sample data of the normal vehicle and performs training;
[0012] S06: Verifying the trained recurrent neural network model using the sample data of the problem vehicle;
[0013] S07: Repeat S05 to S06 until the optimal recurrent neural network model is obtained;
[0014] S08: The optimal recurrent neural network model obtains real-time vehicle data, calculates a prediction result according to a set frequency, compares the prediction result with the actual result to obtain a deviation, and when the deviation is greater than a threshold, it can be determined that the battery has a risk of thermal runaway.
[0015] Preferably,
[0016] In the S01, the historical data includes laboratory data and user's actual vehicle data;
[0017] The characteristic data includes total current, total voltage, SOC, vehicle speed, accumulated mileage, single cell voltage, and temperatures collected by several temperature sensors;
[0018] The method of performing data cleaning on the feature data includes: removing invalid data, filling missing data and normalizing.
[0019] Preferably,
[0020] The formula for normalizing the characteristic data is:
[0021]
[0022] Among them, x represents the original value, Indicates the normalized value.
[0023] Preferably,
[0024] In S02, a total of P1 sample data of fixed time length are generated by sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the feature number.
[0025] Preferably,
[0026] In S03, the recurrent neural network model is composed of a plurality of basic recurrent neural units and a fully connected layer, wherein the basic recurrent neural units include simple recurrent neural units, LSTM units and GRU units;
[0027] The recurrent neural network model may also use a customized recurrent neural unit.
[0028] Preferably,
[0029] In S04, the output of the recurrent neural network model is defined, that is, there are two ways of prediction. First, all data with a length of m1 at a time are predicted based on the acquired data. Second, a cyclic prediction method is adopted, that is, each output is a prediction of the next moment, and then the data with a length of m-P4 in the sample data and the predicted data at the moment with a length of P4 are used to predict the result of the next moment, until all data with a predicted length of m1 at the moment are obtained.
[0030] Preferably,
[0031] In S05, the recurrent neural network model obtains sample data of P2 normal vehicles for training, and the evaluation index of the recurrent neural network model is the mean square error MSE; regularization and batch normalization are adopted to optimize the training process of the recurrent neural network model, and the MSE is defined as parameter MSE 正常 ;
[0032] In S06, the sample data of P3 problem vehicles are used to verify the recurrent neural network model after training, and the evaluation index of the verification result is MSE, which is defined as parameter MSE 问题 ;
[0033] In S07, S05 to S06 are repeated until the MSE is obtained. 正常 Smaller, MSE 问题 Larger optimal recurrent neural network models;
[0034] In S08, the optimal recurrent neural network model obtains real-time vehicle data and calculates the MSE 实时 , the calculation formula of the deviation, i.e. R, is:
[0035]
[0036] The present invention also provides a battery thermal runaway risk prediction system based on a recurrent neural network, comprising:
[0037] A cleaning module, used to select feature data from the historical data collected in advance for training the recurrent neural network model, and perform data cleaning on the feature data;
[0038] A generating module, used for generating sample data of a fixed time length based on the characteristic data after cleaning, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle;
[0039] A construction module, used to construct the recurrent neural network model;
[0040] A definition module, used to define the output of the recurrent neural network model;
[0041] A training module, used for the recurrent neural network model to obtain sample data of the normal vehicle and perform training;
[0042] A verification module, used to verify the trained recurrent neural network model using sample data of the problem vehicle;
[0043] A loop module, used for repeating the training module to the verification module until the optimal recurrent neural network model is obtained;
[0044] The comparison module is used for the optimal recurrent neural network model to obtain real-time vehicle data, calculate the prediction result according to the set frequency, compare the prediction result with the actual result to obtain the deviation, and when the deviation is greater than a threshold, it can be determined that the battery has a thermal runaway risk.
[0045] Preferably,
[0046] In the cleaning module, the historical data includes laboratory data and user's actual vehicle data;
[0047] The characteristic data includes total current, total voltage, SOC, vehicle speed, accumulated mileage, single cell voltage, and temperatures collected by several temperature sensors;
[0048] The method of performing data cleaning on the feature data includes: removing invalid data, filling missing data and normalizing.
[0049] Preferably,
[0050] The formula for normalizing the characteristic data is:
[0051]
[0052] Among them, x represents the original value, Indicates the normalized value.
[0053] Preferably,
[0054] In the generation module, a total of P1 sample data of fixed time length are generated by a sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the number of features.
[0055] Preferably,
[0056] In the construction module, the recurrent neural network model is composed of a plurality of basic recurrent neural units and a fully connected layer, wherein the basic recurrent neural units include simple recurrent neural units, LSTM units and GRU units;
[0057] The recurrent neural network model may also use a customized recurrent neural unit.
[0058] Preferably,
[0059] In the definition module, the output of the defined recurrent neural network model, that is, the prediction method, has two ways: first, predicting all data with a length of m1 at a time based on the acquired data; second, adopting a cyclic prediction method, that is, each output is a prediction for the next moment, and then using the data with a length of m-P4 in the sample data and the predicted data at the moment with a length of P4 to predict the result of the next moment, until all data with a predicted length of m1 at the moment are obtained.
[0060] Preferably,
[0061] In the training module, the recurrent neural network model obtains sample data of P2 normal vehicles for training, and the evaluation index of the recurrent neural network model is the mean square error MSE; regularization and batch normalization are used to optimize the training process of the recurrent neural network model, and the MSE is defined as parameter MSE 正常 ;
[0062] In the verification module, the sample data of P3 problem vehicles are used to verify the trained recurrent neural network model. The evaluation index of the verification result is MSE, which is defined as parameter MSE 问题 ;
[0063] In the loop module, the training module to the validation module are repeated until the MSE is obtained. 正常 Smaller, MSE 问题 Larger optimal recurrent neural network models;
[0064] In the comparison module, the optimal recurrent neural network model obtains real-time vehicle data and calculates the MSE 实时 , the calculation formula of the deviation, i.e. R, is:
[0065]
[0066] By adopting the above technical scheme, the present invention can achieve the following beneficial technical effects: the present invention utilizes the characteristic that vehicle data belongs to time series data, adopts a deep learning method to construct a recurrent neural network model, uses the data of normal vehicles as training data, and uses the data of problem vehicles as verification data, and the characteristics of the data include multiple dimensions such as SOC, current, single cell voltage and temperature; in actual applications, the trained model is deployed to a big data platform, and the real-time data collected is used to perform calculations at a certain frequency, and then the prediction results of the model are obtained, and the prediction results are compared with the actual results to calculate the deviation, which can be used as the probability that the battery may have a thermal runaway failure. The present invention can accurately and timely predict the thermal runaway risk of the battery. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0068] The present invention will be further described below in conjunction with the accompanying drawings.
[0069] like Figure 1 As shown, the present invention provides a battery thermal runaway risk prediction method based on a recurrent neural network, comprising the following steps:
[0070] S01: Select feature data from the historical data collected in advance for training the recurrent neural network model, and perform data cleaning on the feature data.
[0071] Specifically, historical data includes laboratory data and user real vehicle data. Data of vehicles with the same model of power battery are randomly extracted, and the vehicles can be stratified in advance according to mileage. The selected vehicles are all normal vehicles that have not reported any faults. The number of selected vehicles accounts for no less than 10% of the total number of vehicles, and the selected data period is at least one year.
[0072] The characteristic data include total current, total voltage, SOC, vehicle speed, cumulative mileage, single cell voltage, temperature collected by several temperature sensors, etc. In addition, the data of the whole life cycle of the thermal runaway vehicle is collected in the same way.
[0073] The methods for cleaning feature data include: removing invalid data, filling missing data and normalizing.
[0074] Using max-min normalization, the specific formula is:
[0075]
[0076] Among them, x represents the original value, Indicates the normalized value.
[0077] S02: Generate sample data of a fixed time length based on the cleaned feature data, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle.
[0078] Specifically, a total of P1 sample data of a fixed time length is generated by a sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the number of features. The interval of the sliding window should be as small as possible so that as much data as possible can be generated, but it should also meet the requirements of computing power and computing time.
[0079] S03: Build a recurrent neural network model.
[0080] Specifically, the recurrent neural network model consists of several basic recurrent neural units and a fully connected layer. The basic recurrent neural units include simple recurrent neural units, LSTM units, and GRU units. These basic recurrent neural units can be directly called in common machine learning frameworks. The recurrent neural network model can also use custom recurrent neural units.
[0081] S04: Define the output of the recurrent neural network model.
[0082] Specifically, there are two ways to define the output of the recurrent neural network model, that is, to predict the data with a length of m1 at a time based on the acquired data. Second, a cyclic prediction method is used, that is, each output is a prediction for the next moment, and then the data with a length of m-P4 (P4 is 1) in the sample data and the predicted data with a length of P4 (P4 is 1) (the same data with a total length of m) are used to predict the result of the next moment, and so on, until all the data with a predicted length of m1 are obtained. Both methods can give prediction results of any length, and which method is used is determined based on the number of neurons in the fully connected layer.
[0083] S05: The recurrent neural network model obtains sample data of normal vehicles and performs training.
[0084] Specifically, the recurrent neural network model obtains sample data of P1 normal vehicles for training. The evaluation index of the recurrent neural network model is the mean square error MSE. The smaller the MSE, the better. However, in order to avoid overfitting, regularization and batch normalization can be used to optimize the training process of the recurrent neural network model. MSE is defined as parameter MSE 正常 ;
[0085] S06: Verify the trained recurrent neural network model using sample data of the problem vehicle.
[0086] Specifically, the recurrent neural network model after training is verified using sample data of P2 problem vehicles. The evaluation index of the verification result is MSE, which is defined as parameter MSE 问题 , because it is the data of the problem vehicle, MSE 问题 The bigger the better.
[0087] S07: Repeat S05 to S06 until the optimal recurrent neural network model is obtained.
[0088] Specifically, repeat S05 to S06 until the MSE is obtained. 正常 Smaller, MSE 问题 The larger optimal recurrent neural network model is the model that will eventually be put into use.
[0089] S08: The optimal recurrent neural network model obtains real-time vehicle data, calculates the prediction results according to the set frequency, and compares the prediction results with the actual results to obtain the deviation. When the deviation is greater than the threshold, it can be determined that the battery has a thermal runaway risk.
[0090] Specifically, the optimal recurrent neural network model obtains real-time sample data of the vehicle and calculates the MSE 实时 , the calculation formula of deviation, i.e. R, is:
[0091]
[0092] The present invention also provides a battery thermal runaway risk prediction system based on a recurrent neural network, comprising:
[0093] A cleaning module, used to select feature data from the historical data collected in advance for training the recurrent neural network model, and perform data cleaning on the feature data;
[0094] A generating module, used to generate sample data of a fixed time length based on the cleaned feature data, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle;
[0095] Building blocks for building recurrent neural network models;
[0096] Definition module, used to define the output of the recurrent neural network model;
[0097] The training module is used for the recurrent neural network model to obtain sample data of normal vehicles and perform training;
[0098] A verification module is used to verify the trained recurrent neural network model using sample data of problem vehicles;
[0099] The loop module is used to repeat the training module to the verification module until the optimal recurrent neural network model is obtained;
[0100] The comparison module is used to obtain real-time vehicle data using the optimal recurrent neural network model, calculate the prediction results according to the set frequency, and compare the prediction results with the actual results to obtain the deviation. When the deviation is greater than the threshold, it can be determined that the battery has a risk of thermal runaway.
[0101] Specifically,
[0102] In the cleaning module, historical data includes laboratory data and user's actual vehicle data;
[0103] The characteristic data includes total current, total voltage, SOC, vehicle speed, cumulative mileage, single cell voltage, and temperatures collected by several temperature sensors;
[0104] Methods for data cleaning of feature data include: removing invalid data, completing missing data, and normalizing.
[0105] Specifically,
[0106] The formula for normalizing feature data is:
[0107]
[0108] Among them, x represents the original value, Indicates the normalized value.
[0109] Specifically,
[0110] In the generation module, a total of P1 sample data of fixed time length are generated through a sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the number of features.
[0111] Specifically,
[0112] In the building module, the recurrent neural network model consists of several basic recurrent neural units and a fully connected layer. The basic recurrent neural units include simple recurrent neural units, LSTM units and GRU units.
[0113] The recurrent neural network model can also use custom recurrent neural units.
[0114] Specifically,
[0115] In the definition module, the output of the recurrent neural network model is defined, that is, there are two ways of prediction. First, all data with a length of m1 at a time are predicted based on the acquired data. Second, a cyclic prediction method is adopted, that is, each output is a prediction for the next moment, and then the data with a length of m-P4 in the sample data and the predicted data at the moment of length P4 are used to predict the result of the next moment, until all data with a predicted length of m1 at the moment are obtained.
[0116] Specifically,
[0117] In the training module, the recurrent neural network model obtains sample data of P2 normal vehicles for training. The evaluation index of the recurrent neural network model is the mean square error MSE. Regularization and batch normalization are used to optimize the training process of the recurrent neural network model. MSE is defined as parameter MSE 正常 ;
[0118] In the verification module, the sample data of P3 problem vehicles is used to verify the trained recurrent neural network model. The evaluation index of the verification result is MSE, which is defined as parameter MSE 问题 ;
[0119] In the loop module, repeat the training module to the validation module until the MSE is obtained. 正常 Smaller, MSE 问题 Larger optimal recurrent neural network models;
[0120] In the comparison module, the optimal recurrent neural network model obtains real-time vehicle data and calculates the MSE 实时 , the calculation formula of deviation, i.e. R, is:
[0121]
Claims
1. A battery thermal runaway risk prediction method based on recurrent neural network, characterized in that: The following steps are involved: S01: Selecting feature data from historical data collected in advance for training a recurrent neural network model, and performing data cleaning on the feature data; S02: generating sample data of a fixed time length based on the characteristic data after cleaning, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle; S03: construct the recurrent neural network model; S04: defining the output of the recurrent neural network model; S05: The recurrent neural network model obtains sample data of normal vehicles and performs training. The evaluation index of the recurrent neural network model is the mean square error (MSE). Regularization and batch normalization are used to optimize the training process of the recurrent neural network model. The MSE is defined as parameter MSE 正常 ; S06: Validate the trained recurrent neural network model using the sample data of the problem vehicle. The evaluation index of the validation result is MSE, which is defined as parameter MSE 问题 ;; S07: Repeat S05 to S06 until the MSE is obtained. 正常 Smaller, MSE 问题 Larger optimal recurrent neural network models; S08: The optimal recurrent neural network model obtains real-time vehicle data and calculates the prediction result MSE according to the set frequency 实时 , comparing the predicted result with the actual result to obtain a deviation, and when the deviation is greater than a threshold, it can be determined that the battery has a thermal runaway risk; The calculation formula of the deviation, i.e. R, is:
2. The method for predicting battery thermal runaway risk based on a recurrent neural network according to claim 1, characterized in that: In the S01, the historical data includes laboratory data and user's actual vehicle data; The characteristic data includes total current, total voltage, SOC, vehicle speed, accumulated mileage, single cell voltage, and temperatures collected by several temperature sensors; The method of performing data cleaning on the feature data includes: removing invalid data, filling missing data and normalizing.
3. The method for predicting battery thermal runaway risk based on a recurrent neural network according to claim 2, characterized in that: The formula for normalizing the characteristic data is: Among them, x represents the original value, Indicates the normalized value.
4. The method for predicting battery thermal runaway risk based on a recurrent neural network according to claim 1, characterized in that: In S02, a total of P1 sample data of fixed time length are generated by sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the feature number.
5. The method for predicting battery thermal runaway risk based on recurrent neural network according to claim 1, characterized in that: In S03, the recurrent neural network model is composed of a plurality of basic recurrent neural units and a fully connected layer, wherein the basic recurrent neural units include simple recurrent neural units, LSTM units and GRU units; The recurrent neural network model may also use a customized recurrent neural unit.
6. The method for predicting battery thermal runaway risk based on a recurrent neural network according to claim 4, characterized in that: In S04, the output of the recurrent neural network model is defined, that is, there are two ways of prediction. First, all data with a length of m1 at a time are predicted based on the acquired data. Second, a cyclic prediction method is adopted, that is, each output is a prediction of the next moment, and then the data with a length of m-P4 in the sample data and the predicted data at the moment with a length of P4 are used to predict the result of the next moment, until all data with a predicted length of m1 at the moment are obtained.
7. A battery thermal runaway risk prediction system based on recurrent neural network, characterized in that: include: A cleaning module, used to select feature data from the historical data collected in advance for training the recurrent neural network model, and perform data cleaning on the feature data; A generating module, used for generating sample data of a fixed time length based on the characteristic data after cleaning, wherein the sample data includes sample data of a normal vehicle and sample data of a problem vehicle; A construction module, used to construct the recurrent neural network model; A definition module, used to define the output of the recurrent neural network model; The training module is used for the recurrent neural network model to obtain sample data of normal vehicles for training. The evaluation index of the recurrent neural network model is the mean square error (MSE). Regularization and batch normalization are used to optimize the training process of the recurrent neural network model. The MSE is defined as the parameter MSE. 正常 ; A verification module is used to verify the trained recurrent neural network model using the sample data of the problem vehicle. The evaluation index of the verification result is MSE, which is defined as parameter MSE 问题 ; A loop module is used to repeat the training module to the verification module until the MSE is obtained. 正常 Smaller, MSE 问题 Larger optimal recurrent neural network models; Comparison module, used for the optimal recurrent neural network model to obtain real-time vehicle data, and calculate the prediction result MSE according to the set frequency 实时 , comparing the predicted result with the actual result to obtain a deviation, and when the deviation is greater than a threshold, it can be determined that the battery has a thermal runaway risk; The calculation formula of the deviation, i.e. R, is:
8. The battery thermal runaway risk prediction system based on recurrent neural network according to claim 7 is characterized in that: In the cleaning module, the historical data includes laboratory data and user's actual vehicle data; The characteristic data includes total current, total voltage, SOC, vehicle speed, accumulated mileage, single cell voltage, and temperatures collected by several temperature sensors; The method of performing data cleaning on the feature data includes: removing invalid data, filling missing data and normalizing.
9. The battery thermal runaway risk prediction system based on recurrent neural network according to claim 8, characterized in that: The formula for normalizing the characteristic data is: Among them, x represents the original value, Indicates the normalized value.
10. The battery thermal runaway risk prediction system based on recurrent neural network according to claim 7, characterized in that: In the generation module, a total of P1 sample data of fixed time length are generated by a sliding window method, including P2 sample data of normal vehicles and P3 sample data of problem vehicles. The size of each sample data is an m×n matrix, where m is the time length and n is the number of features.
11. The battery thermal runaway risk prediction system based on recurrent neural network according to claim 7, characterized in that: In the construction module, the recurrent neural network model is composed of a plurality of basic recurrent neural units and a fully connected layer, wherein the basic recurrent neural units include simple recurrent neural units, LSTM units and GRU units; The recurrent neural network model may also use a customized recurrent neural unit.
12. The battery thermal runaway risk prediction system based on recurrent neural network according to claim 10, characterized in that: In the definition module, the output of the defined recurrent neural network model, that is, the prediction method, has two ways: first, predicting all data with a length of m1 at a time based on the acquired data; second, adopting a cyclic prediction method, that is, each output is a prediction for the next moment, and then using the data with a length of m-P4 in the sample data and the predicted data at the moment with a length of P4 to predict the result of the next moment, until all data with a predicted length of m1 at the moment are obtained.
