A method for predicting battery thermal runaway based on neural networks

Abstract

The application aims to establish a method for predicting battery thermal runaway based on a neural network, comprising S1: data preparation: based on the massive data accumulated by the laboratory and the user vehicle big data platform, obtaining vehicle battery thermal runaway related data; S2: establishing a feature data set and performing normalization processing; S3: using a convolutional neural network to establish a battery thermal runaway prediction model; S4: training and verifying the prediction model; S5: using the model to identify battery thermal runaway abnormalities. Through the massive data accumulated by the laboratory and the vehicle big data platform, a large number of characteristic parameters are constructed, a neural network deep learning algorithm is used, the law of thermal runaway of the power battery is trained and learned, a model capable of predicting thermal runaway is constructed and generated, and combined with the real-time acquisition of vehicle signal data technology, the risk of thermal runaway of the vehicle can be predicted in real time, and a warning can be issued in time.

Description

Technical Field

[0001] This invention relates to a method for predicting battery thermal runaway, and more specifically to a method for predicting battery thermal runaway based on neural networks. Background Technology

[0002] As the power source for electric vehicles, the health of the battery directly affects the overall vehicle condition. Thermal runaway is one of the most serious problems with power batteries, and once it occurs, it can lead to vehicle fires and, in severe cases, explosions. How to promptly detect vehicles potentially experiencing thermal runaway has always been a crucial issue for major electric vehicle manufacturers. Because batteries gradually degrade over time, undergo their own chemical reactions, and are influenced by varying driver habits and operating conditions, the causes of battery thermal runaway are complex and multifaceted, increasing the difficulty of identifying and predicting such events.

[0003] Current methods for predicting battery thermal runaway primarily analyze and predict its causes from one or more dimensions, such as current, voltage, temperature, individual cell states, and internal battery mechanisms. However, due to the complexity and diversity of the causes of battery thermal runaway, coupled with the environmental conditions of the vehicle, it is difficult to cover all scenarios. These methods are only effective for predicting the effects on a subset of vehicles, while vehicles not covered by these methods still face significant risks.

[0004] Chinese patent document CN109978229A discloses a method for predicting thermal runaway of electric vehicle batteries based on big data machine learning. However, this method only uses battery temperature data and battery fault repair data to build a predictive model for battery thermal runaway. Given the complex and diverse causes of battery thermal runaway, building a model solely based on battery temperature and fault-related data is insufficient to reflect the true causes of thermal runaway, leading to the failure of the prediction model under certain complex conditions.

[0005] Chinese patent document CN112363061A discloses a thermal runaway risk assessment method based on big data. However, the assessment model constructed solely based on battery temperature and the rate of temperature rise is too simplistic and one-sided, leading to a significant discrepancy between the assessed thermal runaway risk and the actual situation. Summary of the Invention

[0006] The purpose of this invention is to establish a method for predicting battery thermal runaway based on neural networks. By using massive amounts of data accumulated from laboratory and vehicle big data platforms, a large number of feature parameters are constructed. Using neural network deep learning algorithms, the thermal runaway patterns of power batteries are trained and learned, and a model that can predict thermal runaway is constructed and generated. Combined with big data real-time acquisition of vehicle signal data technology, the risk of vehicle thermal runaway can be predicted in real time, and timely warnings can be issued.

[0007] To achieve the above objectives, the present invention proposes the following technical solution.

[0008] A method for predicting battery thermal runaway based on neural networks, the method comprising the following steps:

[0009] S1: Data Preparation: Based on the massive amounts of data accumulated from the laboratory and user vehicle big data platform, obtain data related to vehicle battery thermal runaway, including data on normal battery status, data on various abnormal states, data on thermal runaway, and data on vehicle usage behavior.

[0010] S2: Establish the feature dataset and perform normalization. Feature variables include tens of thousands of features such as temperature difference, static pressure difference, individual cell over-temperature, internal resistance difference, self-discharge difference, insulation resistance, and fast charging voltage difference.

[0011] S3: Use convolutional neural networks to build a battery thermal runaway prediction model.

[0012] S4: Training and validating the prediction model.

[0013] S5: Use models to identify battery thermal runaway anomalies.

[0014] Furthermore, step S1 includes the following steps:

[0015] S1-1: Collect data related to battery thermal runaway;

[0016] S1-2: Data validity verification and cleaning.

[0017] Furthermore, step S2 includes the following steps:

[0018] S2-1: Using the collected data, calculate the feature variables used to predict battery thermal runaway and generate a feature dataset;

[0019] S2-2: Check the numerical range of each feature variable and perform feature normalization.

[0020] Furthermore, step S3 includes the following steps:

[0021] S3-1: Divide the dataset according to the proportions to generate training and validation datasets;

[0022] S3-2: Verify the consistency between the training dataset and the validation dataset;

[0023] S3-3: A battery thermal runaway prediction model is established using a convolutional neural network;

[0024] Furthermore, step S4 includes the following steps:

[0025] S4-1: Train the prediction model using the training dataset;

[0026] S4-2: Validate the model accuracy using the validation dataset;

[0027] S4-3: Identify the model with better accuracy and use it as the final prediction model.

[0028] Furthermore, step S5 includes the following steps:

[0029] S5-1: Real-time online collection of vehicle battery-related data;

[0030] S5-2: Calculate the feature variables used as input to the model;

[0031] S5-3: Input the feature variables into the model and output the probability of battery thermal runaway.

[0032] The advantages of this invention are as follows:

[0033] This invention, on the one hand, utilizes massive amounts of data accumulated from laboratories and vehicle big data platforms, including battery temperature data, battery fault repair data, user behavior data, vehicle operating environment data, and battery internal signal data, to establish a predictive model for battery thermal runaway across various diverse and comprehensive dimensions. On the other hand, this method uses massive amounts of data from different dimensions to establish a large number of feature parameters that may be related to thermal runaway, vehicle operating environment, user behavior, and battery internal mechanisms. It leverages the ability of deep learning algorithms in neural networks to effectively learn data patterns from a large number of feature parameters, constructing and generating a convolutional neural network model that can predict power battery thermal runaway. The predictive model trained with massive amounts of data and numerous feature parameters can effectively cover situations where battery thermal runaway occurs. Furthermore, combined with real-time big data stream processing technology, vehicle signal data is collected in real time, and the risk of vehicle thermal runaway is predicted through the convolutional neural network model, issuing timely warnings. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of a method for predicting battery thermal runaway based on neural networks in an embodiment of the present invention; Detailed Implementation

[0035] The present invention will now be further described with reference to the accompanying drawings:

[0036] See Figure 1 This invention is a method for predicting battery thermal runaway based on neural networks, and the specific steps are as follows:

[0037] S1: Data preparation.

[0038] Step S1 specifically includes the following steps:

[0039] S1-1: Collect battery thermal runaway related data, including battery normal state data, data on various abnormal states, data on thermal runaway, vehicle usage behavior, etc.

[0040] In this step, the data related to vehicle battery thermal runaway primarily comes from massive amounts of data accumulated in laboratories and user vehicle big data platforms. Laboratory data is generated after a series of experiments conducted in laboratories under different environmental and operating conditions. User vehicle big data platform data is generated by real vehicle owners during charging and discharging activities under various operating conditions. The collected data fields mainly include total voltage, total current, battery temperature, capacitance, internal resistance, individual cell voltage and current, individual cell temperature, cell voltage, cell temperature, SOC, vehicle speed, charging status, and battery abnormal signal data.

[0041] S1-2: Data validity verification and cleaning.

[0042] Since the collected data often contains noisy data, such as missing values ​​or values ​​outside the normal range, noisy data can cause bias in model predictions. Therefore, the first step is to check and verify whether noisy data exists in the data. If so, the data should be cleaned, for example, by filtering out abnormal data or filling missing data with default values.

[0043] S2: Establish feature datasets and perform normalization.

[0044] Step S2 specifically includes the following steps:

[0045] S2-1: Using the collected battery-related data, calculate the feature variables used to predict battery thermal runaway and generate a feature dataset;

[0046] In this step, the feature variables are used as input variables for the model in subsequent steps: training, validation, and prediction. Based on the data processed in step S1, each feature variable is calculated. The feature variables include tens of thousands of features such as temperature difference, static pressure difference, individual cell over-temperature, internal resistance difference, self-discharge difference, insulation impedance, and fast charging voltage difference.

[0047] S2-2: Check the numerical range of each feature variable and perform feature normalization.

[0048] Different feature variables typically have different value ranges, and these differences can be significant. To minimize model bias and improve model learning efficiency, each feature variable is normalized to bring its values ​​to the range [0,1]. The normalization formula is as follows:

[0049]

[0050] x represents the original value. This represents the normalized value.

[0051] S3: Use convolutional neural networks to build a battery thermal runaway prediction model.

[0052] Step S3 specifically includes the following steps:

[0053] S3-1: Divide the dataset according to the proportions to generate training and validation datasets.

[0054] In this step, the normalized feature dataset from step S2 is divided into a training dataset and a validation dataset in a 7:3 ratio. The training dataset is used to train the model, and the validation dataset is used to test the model's performance.

[0055] S3-2: Verify the consistency between the training dataset and the validation dataset.

[0056] Consistency refers to the consistency between the training and validation datasets in terms of feature types, data distribution, etc. For example, for continuous feature variables, their means, extreme values, medians, etc., should be equal or similar in both the training and validation datasets. For discrete feature variables, their sample values ​​in both datasets should be covered.

[0057] When the training and validation datasets are inconsistent and significantly different, it can lead to biased predictions in the model and reduce its generalization ability.

[0058] S3-3: A battery thermal runaway prediction model is established using a convolutional neural network.

[0059] In this step, a convolutional neural network (CNN) model is used. The number of neurons in the input layer is kept the same as the amount of data in the feature variables. For example, if each sample in the feature dataset has 1000 features, then the input layer will have 1000 neurons.

[0060] Since a convolutional neural network is used, the hidden layers are also called convolutional layers. The model can contain multiple convolutional layers, and each convolutional layer can contain multiple neurons, with each convolutional layer corresponding to a convolutional kernel. The number of convolutional layers λ, the number of neurons in each convolutional layer δ, and the kernel size η are set as hyperparameters. Because the number of neurons and the kernel size in each convolutional layer can be different,

[0061] δ = {δ1, δ2, ... δ} λ}

[0062] η = {η1, η2, ..., η} λ}

[0063] Hyperparameters are not fixed values; they can take different values ​​during iterative model training. When the model performs optimally, the hyperparameter values ​​can be fixed at their current values. The activation function used in convolutional layers is ReLU, which exhibits better learning ability compared to other functions such as the Sigmoid and Tanh functions. The ReLU function formula is as follows:

[0064]

[0065] The output layer has only one neuron, and its activation function is the sigmoid function:

[0066]

[0067] The sigmoid function can output a probability value; the higher the probability, the greater the risk of thermal runaway in the battery.

[0068] The above input layer, hidden layer, and output layer constitute a complete neural network model. In practice, deep learning frameworks such as Tensorflow, Caffe, and PyTorch can be used to build it.

[0069] S4: Training and validating the prediction model.

[0070] Step S4 specifically includes the following steps:

[0071] S4-1: Train the prediction model using the training dataset.

[0072] In this step, the model is trained multiple times using the training dataset until the error rate in the learning curve reaches an inflection point. The hyperparameter values ​​defined in step S3 can be adjusted during training.

[0073] S4-2: Validate the model's performance using the validation dataset.

[0074] Validate the model trained in step S4-1 using the validation dataset. If the AUC value on the validation set is less than ω, it indicates that the model is performing poorly and should be retrained for optimization. The AUC value typically ranges from [0.5, 1], with values ​​closer to 1 indicating better model performance. ω is usually greater than 0.8 and less than 1, and its value can be chosen based on the specific business problem and data performance.

[0075] S4-3: Identify the best-performing models and use them as the final prediction models.

[0076] S5: Use models to identify battery thermal runaway anomalies.

[0077] Step S5 specifically includes the following steps:

[0078] S5-1: Real-time online collection of vehicle battery-related data.

[0079] This step utilizes real-time big data streaming technologies such as Spark Streaming and Flink to collect real-time signal data from the user's vehicle battery.

[0080] S5-2: Calculate the feature variables used to input the model.

[0081] The characteristic variables are calculated from the signal data collected in step S5-1 using the method in step S2.

[0082] S5-3: Input the feature variables into the model and output the probability of battery thermal runaway.

[0083] In this step, a threshold value of э is set for the probability of thermal runaway in the battery. When the probability output by the model is greater than or equal to э, the battery is considered to have experienced thermal runaway and an early warning is issued. Generally, э can be controlled within the range of [0.50, 0.98], and the most suitable value can be obtained by analyzing historical data of actual thermal runaway events.

[0084] As can be seen from the above embodiments, this invention utilizes a convolutional neural network model to establish a large number of feature parameters that may be related to thermal runaway, vehicle operating conditions, user behavior, and battery internal mechanisms, based on massive amounts of data from laboratory and vehicle big data platforms. Leveraging the ability of deep learning algorithms in neural networks to effectively learn data patterns from a large number of feature parameters, a convolutional neural network model capable of predicting thermal runaway of power batteries is constructed and generated. The prediction model trained with massive amounts of data and numerous feature parameters can effectively cover situations where battery thermal runaway occurs.

Claims

1. A method for predicting thermal runaway of a battery based on a neural network, the method comprising: The method comprises the following steps: S1: data preparation: based on the massive data accumulated by the laboratory and the user vehicle big data platform, the vehicle battery thermal runaway related data is obtained, including battery normal state data, various abnormal state data, thermal runaway data, and vehicle behavior data; S2: establish a feature data set and perform normalization processing; the feature variables include temperature difference, static pressure difference, single cell overtemperature, internal resistance difference, self-discharge difference, insulation impedance, and fast charging pressure difference; S3: use a convolutional neural network to establish a battery thermal runaway prediction model; including: S3-1: divide the data set according to the proportion to generate a training data set and a verification data set; S3-2: verify the consistency of the training data set and the verification data set; S3-3: adopt a convolutional neural network to establish a battery thermal runaway prediction model: adopt a convolutional neural network (CNN) model, set the number of convolutional layers λ, the number of neurons δ of each convolutional layer, and the convolution kernel size η as hyperparameters, δ = {δ1, δ2,... δ λ} η = {η1, η2,... η λ} The activation function used in the convolutional layer is ReLu, and the ReLu function formula is as follows: The output layer has only one neuron, and the activation function is a sigmoid function: The probability value can be output through the sigmoid function, and the greater the probability, the greater the risk of battery thermal runaway; S4: train and verify the prediction model, including: S4-1: use the training data set to train the prediction model until the error rate in the learning curve decreases to the inflection point; S4-2: verify the model accuracy using the verification data set, if the AUC value of the model on the verification set is less than ω, it indicates that the model effect is poor, and re-optimization training is performed; the AUC value ranges from 0.5 to 1, and the closer the value to 1, the better the model effect, and ω is valued according to the actual business problem and data performance; S4-3: find the model with better accuracy as the final prediction model; S5: use the model to identify battery thermal runaway anomalies. 2.The method of predicting a thermal runaway of a battery based on a neural network according to claim 1, wherein, The step S1 comprises: S1-1: collect battery thermal runaway related data; the fields of the vehicle battery thermal runaway related data mainly include total voltage, total current, battery temperature, capacitance, internal resistance, single cell voltage and current, single cell temperature, cell voltage, cell temperature, SOC, vehicle speed, charging state, battery abnormal signal data, etc.; S1-2: data validity verification and cleaning. 3.The method of predicting a thermal runaway of a battery based on a neural network according to claim 1, wherein, The step S2 comprises: S2-1: use the collected data to calculate the feature variables for predicting battery thermal runaway, and generate a feature data set; S2-2: check the numerical range of each feature variable and perform feature normalization processing. 4.The method of predicting a thermal runaway of a battery based on a neural network according to claim 1, wherein, The step S5 comprises: S5-1: real-time online collection of vehicle battery related data; S5-2: calculate the feature variables for inputting the model; S5-3: input the feature variables into the model to output the probability of battery thermal runaway.

Images