Battery thermal runaway prediction method, training method and device

Abstract

The invention discloses a battery thermal runaway prediction method, training method and device, and the method comprises the steps: obtaining to-be-predicted battery data, and inputting the to-be-predicted battery data into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model comprises a plurality of sub-models, and the plurality of sub-models are obtained by training based on different machine learning algorithms; performing battery thermal runaway prediction on the to-be-predicted battery data by using the plurality of sub-models to correspondingly obtain a plurality of prediction results; according to the performance indexes of the sub-models and the consistency, calculating to obtain model scores of the sub-models, and according to the model scores, calculating the plurality of prediction results to obtain a final prediction result; wherein the performance indexes are obtained after performance evaluation is carried out on the sub-models, and the consistency is determined according to the difference between the performance indexes of different sub-models. The accuracy and stability of thermal runaway prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the field of battery safety technology, and in particular to a battery thermal runaway prediction method, training method and device. Background Art

[0002] With the increase of global energy pressure and the improvement of environmental awareness, fuel-powered vehicles and large-scale mechanical equipment are gradually being replaced by electric-powered products. New energy vehicles, especially electric vehicles, have gradually become an important development direction of the transportation industry with their advantages of low pollution, low noise, and high efficiency. However, the power battery system of new energy vehicles, especially lithium-ion batteries, is prone to safety problems during the charging and discharging process or in use due to its strong electrochemical activity, the most serious of which is battery thermal runaway.

[0003] Battery thermal runaway refers to the phenomenon that under certain conditions (such as overcharging, over-discharging, short circuit, mechanical abuse, thermal abuse, etc.), the temperature inside the battery rises sharply, accompanied by gas generation and increased pressure, which eventually leads to battery fire or even explosion. This phenomenon not only causes damage to the vehicle, but may also pose a serious threat to the life and property of the driver and surrounding personnel. Therefore, a method that can effectively predict the problem of battery thermal runaway is urgently needed. Summary of the invention

[0004] The present invention provides a battery thermal runaway prediction method, training method and device, aiming to effectively solve the technical problem that the prior art cannot effectively predict battery thermal runaway.

[0005] According to a first aspect of the present invention, the present invention provides a method for predicting thermal runaway of a battery, comprising:

[0006] Obtain battery data to be predicted, and input the battery data to be predicted into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model includes multiple sub-models, and the multiple sub-models are trained based on different machine learning algorithms;

[0007] Use multiple sub-models to predict battery thermal runaway for the battery data to be predicted, and obtain multiple prediction results accordingly;

[0008] The model score of the sub-model is obtained according to the performance index and consistency calculation of the sub-model, and multiple prediction results are calculated according to the model score to obtain the final prediction result; among which, the performance index is obtained after the performance evaluation of the sub-model, and the consistency is determined according to the difference between the performance indicators of different sub-models.

[0009] Furthermore, the model score of the sub-model is obtained according to the performance index and consistency calculation of the sub-model, and the multiple prediction results are calculated according to the model score to obtain the final prediction result, including:

[0010] Obtain multiple performance indicators of the sub-models, and determine the consistency between the sub-models through a cross-validation algorithm based on the performance indicators between different sub-models;

[0011] Obtaining an overall performance index of each sub-model according to the plurality of performance indexes, and determining a first weight for each sub-model according to consistency;

[0012] The model score is obtained according to the total performance index and the first weight calculation, and the model score is used as the second weight of the corresponding prediction result. Voting decisions are made on multiple prediction results to obtain the final prediction result.

[0013] Furthermore, after obtaining the battery data to be predicted, the method further includes:

[0014] Extract multiple features from the battery data to be predicted; the features include statistical features and construction features, the statistical features are obtained by statistical analysis of the battery data to be predicted, and the construction features are obtained by feature construction based on the battery data to be predicted;

[0015] The contribution rate of the feature in each sub-model is calculated according to the feature importance index. The feature importance index is the feature metric in the process of sub-model calculating the feature;

[0016] The comprehensive contribution rate is obtained based on the model score and contribution rate calculation, and the comprehensive contribution rate represents the impact of the characteristics on the final prediction results.

[0017] Furthermore, the machine learning algorithms are various ones of decision tree, random forest, support vector machine, Bayesian, multivariate analysis and neural network.

[0018] Furthermore, when the sub-model is a model based on a decision tree algorithm or a random forest algorithm, the feature importance indicator is information gain;

[0019] When the sub-model is a model based on the Bayesian algorithm, the feature importance indicator is the posterior probability.

[0020] Furthermore, the performance indicators include positive indicators and negative indicators. The positive indicators include one or more of accuracy, precision, recall, and F1 score, and the negative indicators include one or more of mean square error, root mean square error, and mean absolute error.

[0021] Furthermore, the battery data to be predicted includes battery status data, charging condition data, battery traceability data, usage condition data of the product in which the battery is located, and fault data.

[0022] Furthermore, after obtaining the final prediction result, the method further includes:

[0023] When the final prediction result indicates that the current battery has a thermal runaway risk, a battery thermal runaway warning message is generated.

[0024] According to a second aspect of the present invention, the present invention also provides a method for training a battery thermal runaway prediction model, comprising:

[0025] Obtain training battery data and corresponding label information, and perform statistical analysis and feature construction on the training battery data to obtain statistical features and construction features accordingly;

[0026] Target features related to thermal runaway are screened out from statistical features and structural features, and each sub-model is trained based on the target features and label information to obtain a trained sub-model. The sub-model is a model based on different machine learning algorithms. The trained battery thermal runaway prediction model includes multiple trained sub-models.

[0027] Furthermore, the construction features include physical construction features and thermal safety construction features, wherein the physical construction features include one or more of temperature features, power features, and frequency features, and the thermal safety construction features include one or more of heat accumulation features and cumulative temperature features.

[0028] Furthermore, after obtaining the training battery data and the corresponding label information, the method further includes:

[0029] Obtaining a training set and a validation set from the training battery data through a cross-validation algorithm;

[0030] Perform statistical analysis and feature construction on training battery data, including:

[0031] Perform statistical analysis and feature construction on the training set;

[0032] After obtaining the trained sub-model, the method further includes:

[0033] The trained sub-model is validated using the validation set.

[0034] Furthermore, after obtaining the trained sub-model, the method further includes:

[0035] The contribution rate of the target feature in each sub-model is calculated according to the feature importance index. The feature importance index is the feature metric in the process of sub-model calculating the target feature.

[0036] The comprehensive contribution rate is calculated based on the model score and contribution rate, and the comprehensive contribution rate represents the impact of the target feature on the final prediction result.

[0037] According to a third aspect of the present invention, the present invention further provides a battery thermal runaway prediction device, comprising:

[0038] A first data acquisition module is used to acquire battery data to be predicted, and input the battery data to be predicted into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model includes multiple sub-models, and the multiple sub-models are trained based on different machine learning algorithms;

[0039] A prediction module, used to use multiple sub-models to respectively predict battery thermal runaway for the battery data to be predicted, and obtain multiple prediction results accordingly;

[0040] The fusion module is used to obtain the model score of the sub-model according to the performance index and consistency calculation of the sub-model, and calculate multiple prediction results according to the model score to obtain the final prediction result; among which, the performance index is obtained after the performance evaluation of the sub-model, and the consistency is determined according to the difference between the performance indicators of different sub-models.

[0041] According to a fourth aspect of the present invention, the present invention further provides a training device for a battery thermal runaway prediction model, comprising:

[0042] The second data acquisition module is used to acquire training battery data and corresponding label information, and perform statistical analysis and feature construction on the training battery data to obtain statistical features and construction features accordingly;

[0043] The training module is used to screen out target features related to thermal runaway from statistical features and structural features, and train each sub-model based on the target features and label information to obtain a trained sub-model. The sub-model is a model based on different machine learning algorithms. The trained battery thermal runaway prediction model includes multiple trained sub-models.

[0044] According to a fifth aspect of the present invention, the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0045] According to a sixth aspect of the present invention, the present invention further provides a storage medium, in which a plurality of instructions are stored, and the instructions are suitable for being loaded by a processor to execute the steps of the above-mentioned method.

[0046] According to a seventh aspect of the present invention, the present invention further provides a computer program product, comprising a computer program, which implements the steps of the above method when executed by a processor.

[0047] Through one or more of the above embodiments of the present invention, at least the following technical effects can be achieved:

[0048] In the technical solution disclosed in the present invention, the battery thermal runaway prediction method predicts battery thermal runaway for the battery data to be predicted through multiple sub-models, and obtains multiple prediction results accordingly; the multiple prediction results are calculated according to the model scores of the sub-models, so as to obtain the final prediction result. Among them, the multiple sub-models are trained based on different machine learning algorithms, and the model scores are determined based on the performance indicators of the sub-models and the consistency between the sub-models. The better the performance of a single sub-model and the better the consistency with other sub-models, the higher the corresponding model score, the more reliable the prediction result with the model score, and the greater the impact on the final prediction result, thereby improving the accuracy and stability of the final prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The technical solutions and other beneficial effects of the present invention will be made apparent by describing in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0050] Figure 1 A flowchart of a battery thermal runaway prediction method provided by an embodiment of the present invention;

[0051] Figure 2 for Figure 1 The sub-step flow chart of S103;

[0052] Figure 3 A flowchart of the steps of comprehensive contribution calculation provided by an embodiment of the present invention;

[0053] Figure 4 A flowchart of a method for training a battery thermal runaway prediction model provided by an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of the process of training and predicting a battery thermal runaway prediction model provided by an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of the structure of a battery thermal runaway prediction device provided by an embodiment of the present invention;

[0056] Figure 7 A schematic diagram of the structure of a battery thermal runaway prediction model training device provided by an embodiment of the present invention;

[0057] Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The technical scheme in the embodiment of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0059] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the term "and / or" herein is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " herein, unless otherwise specified, generally indicates that the associated objects before and after are in an "or" relationship.

[0060] See also Figure 1 and Figure 5 The battery thermal runaway prediction method provided by the embodiment of the present invention comprises the following steps:

[0061] S101, obtaining battery data to be predicted, and inputting the battery data to be predicted into a pre-trained battery thermal runaway prediction model.

[0062] The battery thermal runaway prediction model includes multiple sub-models. Different sub-models use different machine learning algorithms. Each sub-model is trained using the same training battery data. Each sub-model can predict battery thermal runaway. The machine learning algorithm can be a decision tree, random forest, support vector machine, Bayesian, multivariate analysis, neural network, etc., without limitation.

[0063] The battery data to be predicted includes data such as battery status, charging conditions, battery manufacturing, etc., and may also include data of products equipped with the battery. The data of the product especially includes data related to the battery, such as the usage of the product, the battery temperature during use, and data related to the battery when the product fails. Products equipped with batteries can be electric vehicles, drones, mobile phones, laptops, power tools, medical equipment, etc. Data such as battery status, charging conditions, battery manufacturing, etc. can provide more direct information for battery thermal runaway analysis, and product data can provide indirect information. Data such as battery status, charging conditions, battery manufacturing, etc. and product data can provide diverse and complete data for battery thermal runaway prediction to improve the accuracy of battery thermal runaway prediction. Correspondingly, the training battery data includes data such as battery status, charging conditions, battery manufacturing, etc. and product data. Based on the training battery data, a sub-model with more and more complete model parameters can be trained so that the sub-model can mine features related to thermal runaway in a variety of data, thereby improving the prediction results.

[0064] Battery status data includes electrochemical parameters (voltage, current, capacity, etc.), current temperature data (surface temperature, internal temperature, etc.), etc. Charging condition data includes ambient temperature during charging, battery temperature during charging, number of fast and slow charges, etc. Battery manufacturing data includes battery model, materials, etc. Taking electric vehicles as an example of products equipped with batteries, the corresponding product data includes vehicle driving condition data and vehicle fault data, among which vehicle driving condition data includes vehicle mileage, output energy, battery temperature during driving, etc., and vehicle fault data includes fault data caused by batteries, etc.

[0065] It should be noted that the collected battery data to be predicted may only include battery status, charging conditions, battery manufacturing data and part of product data, and the training battery data includes the above different data. The battery can be a lithium battery, a nickel-cadmium battery, a nickel-hydrogen battery, a lead-acid battery, etc., without limitation.

[0066] S102, using multiple sub-models to perform battery thermal runaway prediction on the battery data to be predicted, and correspondingly obtaining multiple prediction results.

[0067] Among them, the prediction result can be the thermal runaway risk type corresponding to the current battery, for example, the current battery has a thermal runaway risk or does not have a thermal runaway risk, and for another example, the thermal runaway risk of the current battery can be high risk, low risk or no risk, etc. Different thermal runaway risk levels are divided according to probability thresholds. The prediction result can also be the maximum probability value of the thermal runaway risk type to which the current battery belongs. If a sub-model predicts that the probability of the thermal runaway risk type of the current battery being high risk is 0.9, the probability of low risk is 0.19, and the probability of no risk is 0.01, then the prediction result is the maximum probability value of 0.9.

[0068] S103, obtaining a model score of the sub-model according to the performance index and consistency calculation of the sub-model, and calculating multiple prediction results according to the model score to obtain a final prediction result.

[0069] Among them, the performance index is obtained by evaluating the performance of the sub-model after the training of the sub-model is completed. The performance indicators are specifically divided into positive indicators and negative indicators. Positive indicators can be accuracy, precision, recall, F1 score, AUC value (Area Under the ROC Curve), etc., and negative indicators include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), etc. The larger the value of the positive indicator, the better the performance of the sub-model; the smaller the value of the negative indicator, the better the performance of the sub-model. Each sub-model corresponds to multiple performance indicators. The number of performance indicators of each sub-model can be the same or different, and there is no limitation on this.

[0070] Consistency is determined based on the consistency between the performance indicators of different sub-models, that is, the difference between the same performance indicators of different sub-models. The stability and reliability of different sub-models can be determined by calculating and analyzing the mean value and standard deviation of the performance indicators. The smaller the difference between the performance indicators, the higher the consistency. Each sub-model has good stability, generalization ability, reliability and robustness.

[0071] The larger the value of the positive index of the sub-model or the smaller the value of the negative index, the higher the consistency between the sub-models, the higher the model score; conversely, the smaller the value of the positive index of the sub-model or the larger the value of the negative index, the lower the consistency between the sub-models, the lower the model score. Each sub-model corresponds to a model score.

[0072] In this step, the model score can be used as the weight of the prediction result, and the final prediction result can be calculated by weighted summing of multiple prediction results. It can also be based on the model score, and the final prediction result is determined from multiple prediction results through an integrated learning algorithm, where the integrated learning algorithm includes an aggregation algorithm, a stacking algorithm, a voting classification algorithm, etc., which is not limited to this. Among them, the higher the model score, the higher the reliability of the corresponding prediction result, and the greater the impact on the final prediction result.

[0073] The battery thermal runaway prediction method provided by the embodiment of the present invention predicts battery thermal runaway for the battery data to be predicted through multiple sub-models, and obtains multiple prediction results accordingly; the multiple prediction results are calculated according to the model scores of the sub-models, so as to obtain the final prediction result. Among them, the multiple sub-models are trained based on different machine learning algorithms, and the model scores are determined based on the performance indicators of the sub-models and the consistency between the sub-models. The better the performance of a single sub-model and the better the consistency with other sub-models, the higher the corresponding model score, the more reliable the prediction result with the model score, and the greater the impact on the final prediction result, thereby improving the accuracy and stability of the final prediction result.

[0074] In some embodiments of the present invention, Figure 2 As shown, S103 includes the following sub-steps:

[0075] S1031, obtaining multiple performance indicators of the sub-models, and determining the consistency between the sub-models through a cross-validation algorithm based on the performance indicators between different sub-models. The cross-validation algorithm may be K-fold cross-validation, stratified K-fold cross-validation, and the like.

[0076] S1032, obtaining a total performance index of each sub-model according to the calculation of the multiple performance indexes, and determining a first weight for each sub-model according to the consistency.

[0077] In this step, the total performance index can be calculated by calculating the average value of multiple performance indicators, weighted summation, etc. Preferably, the performance indicators of each sub-model are normalized in units of sub-models, and the normalized performance indicators are accumulated to obtain the total performance index, which is specifically as follows:

[0078]

[0079] or,

[0080]

[0081] Where, Value ij is the jth performance index of the ith sub-model, Min_Value is the minimum performance index among the total J performance indexes of the ith sub-model, Max_Value is the maximum performance index among the total J performance indexes of the ith sub-model, n ij is the normalized value of the jth performance index of the ith sub-model, N i is the overall performance index of the ith sub-model, with a total of I sub-models.

[0082] The first weight is set according to the consistency. When the consistency between the current sub-model and other sub-models is high, a higher first weight is set for the current sub-model. Conversely, when the consistency between the current sub-model and other sub-models is low, a lower weight is set for the current sub-model. Schematically, when the difference between the current sub-model and other sub-models is less than the first threshold, the first weight value corresponding to the first threshold is configured for the current sub-model; when the difference is greater than or equal to the first threshold and less than the second threshold, the first weight value corresponding to the second threshold is configured for the current sub-model, and so on. As the difference increases, the consistency becomes worse, and the corresponding first weight value becomes smaller. Alternatively, the first weight value corresponding to the current sub-model is determined based on the mapping relationship between consistency and the first weight (for example, a mapping relationship table is formed by obtaining multiple sets of consistency and first weights through multiple experiments), and so on.

[0083] In other embodiments of the present application, in addition to considering consistency, the setting of the first weight can also be set based on expert experience and domain knowledge. For example, expert experience determines that the decision tree algorithm in the machine learning algorithm has a better thermal runaway risk prediction effect for temperature data in battery data, then a higher first weight is set for the sub-model using the decision tree algorithm based on consistency and expert experience.

[0084] S1033, obtaining a model score according to the total performance index and the first weight calculation, and using the model score as a second weight corresponding to the prediction result, and performing a voting decision on the multiple prediction results to obtain a final prediction result.

[0085] In this step, the product of the first weight and the total performance index is used as the model score of the sub-model: Score i is the model score of the ith sub-model, w i is the first weight of the i-th sub-model. Weighted voting is performed based on the product of the second weight and the predicted result: Vote k is the vote rate of the kth thermal runaway risk type, I(·) is the indicator function, which returns 1 if the condition in the brackets is met, otherwise it returns 0, y ik is the prediction result of the ith sub-model, that is, the thermal runaway risk type to which the current battery belongs. There are K thermal runaway risk types in total. Assuming that the thermal runaway risk types include risky and non-risky, the corresponding risky type is represented by 1 and the non-risky type is represented by 0. When calculating the risky vote rate, if the prediction result of the ith sub-model is risky, then I(y ik =k) ​​= 1, if the prediction result is risk-free, then I(y ik=k)=0. Finally, the prediction result with the highest vote rate is output as the final prediction result. In other embodiments of the present application, the thermal runaway risk types may be divided into three, four, five, etc. for voting decisions, and the present application does not limit this.

[0086] The battery thermal runaway prediction method provided in an embodiment of the present invention obtains a model score of the sub-model based on the performance index of the sub-model and the consistency between the sub-model and other sub-models, and performs voting decisions on multiple prediction results based on the model score. The higher the model score, the greater the influence of its corresponding prediction result on the determination of the final prediction result, thereby improving the classification accuracy and stability.

[0087] In some embodiments of the present invention, after obtaining the battery data to be predicted, the battery thermal runaway prediction method further includes: Figure 3 Steps shown:

[0088] S201, extracting multiple features from battery data to be predicted.

[0089] In this step, after obtaining the original battery data to be predicted, features can be extracted from more battery data to be predicted through statistical analysis, feature engineering and other methods. The features specifically include statistical features and structural features. Statistical features can be obtained by statistical analysis methods such as mean, variance, median, standard deviation, etc. for different types of battery temperatures (including battery temperature during vehicle driving, battery temperature during charging, etc.). Structural features are new feature quantities constructed based on physical principles and battery thermal safety theory, which can specifically be battery aging index, heat accumulation, cumulative temperature impact, etc. The battery data mapped by different features may be the same or different.

[0090] S202, calculating the contribution rate of the feature in each sub-model according to the feature importance index.

[0091] Among them, the feature importance index is the feature metric used by the sub-model to calculate the feature. Different sub-models have different feature importance indicators. For example, in the sub-model of the decision tree algorithm, the feature importance index can be information gain, Gini gain, number of splits, etc.; in the sub-model of the Bayesian algorithm, the feature importance index can be the posterior probability; in the sub-model of the neural network, the feature importance index can be weight, sensitivity, etc. The higher the contribution rate, the greater the impact of the feature on the prediction result, the more important it is, and the more important the data related to the feature is, which is of great significance for the subsequent analysis of the cause of battery thermal runaway.

[0092] The contribution rate of the feature to the prediction result output by the sub-model is determined by calculating and analyzing the feature importance index. Schematically, in the sub-model of the decision tree algorithm, since the splitting of each node in the decision tree is based on certain rules during the construction process, the rule makes the sub-nodes more pure. In this process, the contribution rate of the feature to the sub-model can reflect the role of the feature in the above splitting process. Specifically, the information gain for the battery data D to be predicted can be calculated first: k is the kth thermal runaway risk type, p i is the probability of the kth class in the data set D. The information gain Gain(D,A) of feature A is:

[0093]

[0094] In the formula, v represents the different values ​​of feature A, D m It means that the value of feature A divides the data D into the mth subset among v subsets.

[0095] Feature A has different information gains at different nodes in different decision trees. The contribution rate of feature A in the sub-model is obtained by calculating the average or weighted average of the information gains of all nodes.

[0096] For example, in the sub-model of random forest, since random forest is composed of multiple decision trees, the contribution rate of feature A in the sub-model can also be calculated by information gain, Gini gain, number of splits, etc. Specifically, it can be calculated by the average value of information gain or Gini gain of all decision trees, or it can be determined according to the number of splits of each decision tree.

[0097] In the sub-model of the neural network, specifically, when feature A is input into the sub-model for calculation, the change in the prediction result of the sub-model is observed by changing the numerical size of feature A, and the contribution rate of feature A is determined through this sensitivity analysis.

[0098] In the sub-model of the Bayesian algorithm, for each feature, its conditional probability under different categories is calculated. If the difference between the two conditional probabilities is large, it means that the feature plays a greater role in distinguishing different categories and its contribution rate is higher. The contribution rate is calculated based on the difference.

[0099] S203, obtaining a comprehensive contribution rate according to the model score and the contribution rate calculation, wherein the comprehensive contribution rate represents the influence of the feature on the final prediction result.

[0100] In this step, the comprehensive contribution rate of the feature to the final prediction result is calculated based on the performance of the sub-model, the consistency with other sub-models, and the contribution rate of the feature to the sub-model. When the model score is high and the contribution rate of the feature in each sub-model is high, the comprehensive contribution rate of the feature is also high, and its impact on the final prediction result is greater. The more important the data corresponding to the feature is, the easier it is to analyze and study in the future. Schematically, if the contribution rate of feature A in sub-models 1, 2, and 3 is 0.3, 0.25, and 0.2 respectively, and the model scores of sub-models 1, 2, and 3 are 0.4, 0.4, and 0.2 respectively, then the comprehensive contribution rate = 0.3*0.4+0.25*0.4+0.2*0.2 = 0.26.

[0101] The battery thermal runaway prediction method provided in an embodiment of the present invention calculates the contribution rate of different features in the sub-model through the feature importance index, and calculates the comprehensive contribution rate of the feature according to the model score and contribution rate of the sub-model. The higher the comprehensive contribution rate, the greater the influence of the corresponding feature on the final prediction result and the more important it is, so that the relevant features are visualized in the thermal runaway risk prediction process, which is convenient for analyzing the data of the relevant features.

[0102] In some embodiments of the present invention, after S103, the battery thermal runaway prediction method also includes generating battery thermal runaway warning information when the final prediction result is characterized as the current battery having a thermal runaway risk. In one embodiment, if the prediction result only has two situations: the presence of a thermal runaway risk and the absence of a thermal runaway risk, then when the final prediction result shows that there is a thermal runaway risk, a battery thermal runaway warning information is generated to remind relevant operation and maintenance personnel to handle the current battery in a timely manner. In another embodiment, if the prediction results include four situations: high risk, medium risk, low risk, and no risk, when the final prediction result shows high risk, medium risk, or low risk, a battery thermal runaway warning information is generated, but the urgency or warning method of the battery thermal runaway warning information corresponding to different risk levels is different. For example, when the risk is low, it can only be a pop-up window prompt, and the medium risk is a combination of pop-up windows and voice broadcasts. The high risk can be warned through a remote phone call. In other embodiments of the present application, the types of prediction results may also be 3, 5, etc., and different warning methods may be used according to the risk level to remind operation and maintenance personnel to promptly handle batteries with thermal runaway risks to ensure the safety of products equipped with the batteries.

[0103] The present invention also provides a method for training a battery thermal runaway prediction model, such as Figure 4 as well as Figure 5 As shown, the training method of the battery thermal runaway prediction model includes the following steps:

[0104] S301, obtaining training battery data and corresponding label information, and performing statistical analysis and feature construction on the training battery data to obtain corresponding statistical features and construction features.

[0105] Among them, the training battery data is the relevant data of multiple batteries, and each battery corresponds to a label information. The label information shows whether the battery has a thermal runaway risk, or the thermal runaway risk level. The relevant data of the battery specifically includes battery status data, charging condition data, battery traceability data, and the usage condition data and fault data of the product where the battery is located. Taking the vehicle as an example, the battery status data includes the battery status description SOX (State Of X), the number of relay closures, power on and off conditions, driving time, and static time; driving condition data includes mileage, total output energy, average energy consumption per 100 kilometers, driving environment temperature, driving battery temperature, etc.; charging condition data includes charging environment temperature, charging battery temperature, fast and slow charging amount, fast and slow charging times, fast and slow charging current, etc.; various types of vehicle fault data, especially battery pack related fault data; battery pack traceability production and manufacturing data, etc. Through the above-mentioned richer and more diverse data, a better performance battery thermal runaway prediction model can be trained. The above-mentioned various training battery data will be stored in the data center.

[0106] After obtaining the above-mentioned large amount of data, statistical features can be extracted from the training battery data through statistical analysis methods, and new structural features can be extracted from the training battery data through feature engineering methods.

[0107] S302, selecting target features related to thermal runaway from statistical features and structural features, and training each sub-model based on the target features and label information to obtain a trained sub-model. The trained battery thermal runaway prediction model includes multiple trained sub-models.

[0108] In this step, the target features related to thermal runaway can be screened out from the statistical features and structural features by methods such as correlation analysis, recursive feature elimination, and feature selection based on sub-models, and the target features are input into each sub-model, and the sub-model is trained according to the label information until the preset number of iterations is reached or the performance of the sub-model reaches a certain performance threshold or other training termination conditions, thereby obtaining a trained sub-model. The trained battery thermal runaway prediction model includes the aforementioned trained sub-models.

[0109] The training method of the battery thermal runaway prediction model provided by the embodiment of the present invention trains the sub-model through a large amount of different types of training battery data, thereby improving the performance of the sub-model. At the same time, the dimension of the training battery data is reduced through feature extraction and screening, the calculation complexity is reduced, and the prediction efficiency is improved.

[0110] In addition, the above statistical features can be obtained by calculating the mean, variance, deviation, kurtosis, etc. of the collected training battery data, and the structural features can be constructed according to physical principles and thermal safety theories.

[0111] Among them, the structural characteristics based on physical principles include temperature-related characteristics, energy and power-related characteristics, operating frequency characteristics, etc. The structural characteristics based on thermal safety theory include heat accumulation, cumulative temperature influence, etc.

[0112] Temperature-related features include temperature variation, temperature change rate, high and low temperature duration, etc. The temperature variation can specifically be the range of battery temperature changes under different operating conditions, such as the maximum battery temperature variation during driving and the maximum battery temperature variation during charging. The temperature variation rate is the battery temperature variation rate calculated under different operating conditions, such as the temperature variation rate during driving and the temperature variation rate during charging. The high and low temperature duration is the duration that the battery temperature is above or below a certain threshold, which is calculated under different operating conditions such as driving and charging.

[0113] Energy and power-related characteristics include charging power and discharging power, high-power duration and ratio, etc. The charging power and discharging power can specifically be the average charging power during charging and the average discharging power during driving. The power conditions under different working conditions such as fast charging and slow charging, rapid acceleration and rapid deceleration can also be calculated separately. The high-power duration is the duration of power above a certain threshold (such as a high-rate fast charging power threshold) under different working conditions such as charging or driving.

[0114] The operating frequency characteristics include charging and discharging frequency, high and low temperature environment frequency, etc.

[0115] Heat accumulation is the heat accumulated by the heat generated by the internal resistance of the battery, which can be calculated based on the current, time, internal resistance, etc. under different operating conditions.

[0116] Cumulative temperature impact refers to the cumulative temperature impact over time, which can be obtained by integrating the temperature impact factor at each time step.

[0117] The training method of the battery thermal runaway prediction model provided by the embodiment of the present invention obtains diverse and complete features from a large amount of training battery data through statistical analysis and feature engineering, which provides rich materials for the training of the battery thermal runaway prediction model, thereby improving the accuracy of the prediction.

[0118] In some embodiments of the present invention, in the process of battery thermal runaway prediction model, training is performed through a cross-validation algorithm, specifically: a training set and a validation set are obtained from the training battery data through a cross-validation algorithm, statistical analysis and feature construction are performed on the training set, and the trained sub-model is verified using the validation set. In different training rounds, the division of the training set and the validation set is different, and when the validation set verifies the sub-model, a cross-validation consistency calculation is also performed, and the training of the sub-model is continuously optimized based on the cross-validation consistency, so that the prediction results output by each sub-model have a high consistency.

[0119] The battery thermal runaway prediction model training method provided in the embodiment of the present invention improves the consistency between sub-models through multiple cross-validation, thereby ensuring the consistency of multiple prediction results and improving the accuracy of the final prediction result.

[0120] In addition, the battery data for training includes a test set in addition to a training set and a validation set. After the sub-model is trained, the sub-model is tested for performance using the test set. The validation set is input into the sub-model after each training round of the sub-model. The parameters in the sub-model are adjusted and optimized based on the prediction results corresponding to the validation set. During the sub-model optimization process, the hyperparameters of the sub-model can be adjusted using the Bayesian optimization algorithm to find the best hyperparameter combination. The L1 / L2 regularization strength is set through regularization to constrain the complexity of the model, prevent overfitting, and improve the robustness of the model.

[0121] In some embodiments of the present invention, after obtaining the trained sub-model, the method further includes:

[0122] The contribution rate of each target feature in each sub-model is calculated based on the feature importance index, which is a key indicator in the process of battery thermal runaway prediction based on the target feature.

[0123] Based on the model score and contribution rate calculation of the sub-model, the comprehensive contribution rate of each target feature to the prediction result is obtained. The comprehensive contribution rate characterizes the impact of the target feature on the prediction result. The calculation of the comprehensive contribution rate is the same as that in the above-mentioned battery thermal runaway prediction method and will not be repeated here. During the training stage, the important data is determined from the massive training battery data based on the data corresponding to the feature through the calculation of the comprehensive contribution rate, which provides a reference for subsequent thermal runaway analysis or the collection of key data.

[0124] In some embodiments of the present invention, during the training and inference process of the battery thermal runaway prediction model, before feature extraction of the data, the data will be preprocessed by cleaning, denoising, outlier removal, deduplication, standardization, and formatting to make the preprocessed data more suitable for the calculation of the sub-model.

[0125] The embodiment of the present invention also provides a battery thermal runaway prediction device, such as Figure 6 As shown, the battery thermal runaway prediction device includes:

[0126] The first data acquisition module 601 is used to acquire the battery data to be predicted, and input the battery data to be predicted into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model includes multiple sub-models, and the multiple sub-models are trained based on different machine learning algorithms.

[0127] The prediction module 602 is used to use multiple sub-models to perform battery thermal runaway prediction on the battery data to be predicted, and obtain multiple prediction results accordingly.

[0128] The fusion module 603 is used to obtain the model score of the sub-model according to the performance index and consistency calculation of the sub-model, and to calculate multiple prediction results according to the model score to obtain the final prediction result; wherein, the performance index is obtained after the performance evaluation of the sub-model, and the consistency is determined according to the difference between the performance indicators of different sub-models.

[0129] The battery thermal runaway prediction device in this embodiment corresponds to the above-mentioned battery thermal runaway prediction method, which will not be described in detail here.

[0130] The embodiment of the present invention also provides a training device for a battery thermal runaway prediction model, such as Figure 7 As shown, the training device of the battery thermal runaway prediction model includes:

[0131] The second data acquisition module 701 is used to acquire training battery data and corresponding label information, and perform statistical analysis and feature construction on the training battery data to obtain statistical features and construction features accordingly.

[0132] The training module 702 is used to filter out target features related to thermal runaway from statistical features and structural features, and train each sub-model based on the target features and label information to obtain a trained sub-model. The sub-model is a model based on different machine learning algorithms. The trained battery thermal runaway prediction model includes multiple trained sub-models.

[0133] The training device for the battery thermal runaway prediction model in this embodiment corresponds to the training method for the battery thermal runaway prediction model described above, and will not be described in detail here.

[0134] Based on any of the above embodiments, another embodiment of the present application further provides an electronic device, such as Figure 8As shown, the electronic device may include: a processor 810 (Processor), a communication interface 820 (Communications Interface), a memory 830 (Memory) and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor may call the logic instructions in the memory 830 to execute the above-mentioned battery thermal runaway prediction method or the battery thermal runaway prediction model training method.

[0135] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0136] On the other hand, an embodiment of the present application further provides a storage medium on which a plurality of instructions are stored, and the instructions are suitable for being loaded by a processor to execute a battery thermal runaway prediction method or a battery thermal runaway prediction model training method as provided in the above embodiments.

[0137] On the other hand, an embodiment of the present application further provides a computer program product. When the program code contained in the computer program product is executed by a processor in an electronic device, the battery thermal runaway prediction method or the battery thermal runaway prediction model training method provided in the above embodiments is implemented.

[0138] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0139] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0140] In summary, although the present invention has been disclosed as above in terms of preferred embodiments, the above preferred embodiments are not intended to limit the present invention. A person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined in the claims.

Claims

1. A method for predicting battery thermal runaway, characterized in that: include: Acquire battery data to be predicted, and input the battery data to be predicted into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model includes a plurality of sub-models, and the plurality of sub-models are trained based on different machine learning algorithms; Using the multiple sub-models to respectively perform battery thermal runaway prediction on the battery data to be predicted, and correspondingly obtain multiple prediction results; A model score of the sub-model is obtained according to the performance index and consistency calculation of the sub-model, and the multiple prediction results are calculated according to the model score to obtain a final prediction result; wherein the performance index is obtained after performance evaluation of the sub-model, and the consistency is determined according to the difference between the performance indexes of different sub-models.

2. The method according to claim 1, characterized in that The step of obtaining a model score of the sub-model according to the performance index and consistency calculation of the sub-model, and calculating the multiple prediction results according to the model score to obtain a final prediction result includes: Acquire multiple performance indicators of the sub-models, and determine the consistency between the sub-models through a cross-validation algorithm based on the performance indicators between different sub-models; Calculating an overall performance index of each sub-model according to the multiple performance indexes, and determining a first weight for each sub-model according to the consistency; A model score is obtained according to the total performance index and the first weight calculation, and the model score is used as the second weight of the corresponding prediction result, and a voting decision is made on the multiple prediction results to obtain a final prediction result.

3. The method according to claim 1, characterized in that After obtaining the battery data to be predicted, the method further includes: Extracting multiple features from the battery data to be predicted; the features include statistical features and construction features, the statistical features are obtained by statistically analyzing the battery data to be predicted, and the construction features are obtained by constructing features based on the battery data to be predicted; Calculating the contribution rate of the feature in each sub-model according to the feature importance index, wherein the feature importance index is a feature metric in the process of calculating the feature by the sub-model; A comprehensive contribution rate is calculated based on the model score and the contribution rate, and the comprehensive contribution rate represents the influence of the feature on the final prediction result.

4. The method according to claim 3, characterized in that The machine learning algorithms are multiple ones of decision tree, random forest, support vector machine, Bayesian, multivariate analysis and neural network.

5. The method according to claim 4, characterized in that When the sub-model is a model based on a decision tree algorithm or a random forest algorithm, the feature importance indicator is information gain; When the sub-model is a model based on the Bayesian algorithm, the feature importance indicator is the posterior probability.

6. The method according to claim 1, characterized in that The performance indicators include positive indicators and negative indicators. The positive indicators include one or more of accuracy, precision, recall rate, and F1 score. The negative indicators include one or more of mean square error, root mean square error, and mean absolute error.

7. The method according to any one of claims 1 to 6, characterized in that: The battery data to be predicted includes battery status data, charging condition data, battery traceability data, usage condition data of the product in which the battery is located, and fault data.

8. The method according to claim 7, characterized in that After obtaining the final prediction result, the method further includes: When the final prediction result indicates that the current battery has a thermal runaway risk, battery thermal runaway warning information is generated.

9. A method for training a battery thermal runaway prediction model, characterized in that: include: Acquire training battery data and corresponding label information, and perform statistical analysis and feature construction on the training battery data to obtain statistical features and construction features accordingly; Target features related to thermal runaway are screened out from the statistical features and the structural features, and each sub-model is trained based on the target features and the label information to obtain a trained sub-model, wherein the sub-model is a model based on different machine learning algorithms, and the trained battery thermal runaway prediction model includes multiple trained sub-models.

10. The method according to claim 9, characterized in that The structural features include physical structural features and thermal safety structural features, wherein the physical structural features include one or more of temperature features, power features, and frequency features, and the thermal safety structural features include one or more of heat accumulation features and cumulative temperature features.

11. The method according to claim 9, characterized in that After obtaining the training battery data and the corresponding label information, the method further includes: Obtaining a training set and a validation set from the training battery data by a cross-validation algorithm; The statistical analysis and feature construction of the training battery data includes: Performing statistical analysis and feature construction on the training set; After obtaining the trained sub-model, the method further includes: The trained sub-model is verified using the verification set.

12. The method according to claim 9, characterized in that After obtaining the trained sub-model, the method further includes: Calculating the contribution rate of the target feature in each sub-model according to the feature importance index, wherein the feature importance index is a feature metric in the process of calculating the target feature by the sub-model; The comprehensive contribution rate is calculated based on the model score and the contribution rate, and the comprehensive contribution rate represents the influence of the target feature on the final prediction result.

13. A battery thermal runaway prediction device, characterized in that: include: A first data acquisition module is used to acquire battery data to be predicted, and input the battery data to be predicted into a pre-trained battery thermal runaway prediction model; the battery thermal runaway prediction model includes a plurality of sub-models, and the plurality of sub-models are trained based on different machine learning algorithms; A prediction module, used to use the multiple sub-models to respectively predict battery thermal runaway for the battery data to be predicted, and obtain multiple prediction results accordingly; A fusion module is used to obtain a model score of a sub-model based on the performance index and consistency calculation of the sub-model, and to calculate the multiple prediction results based on the model score to obtain a final prediction result; wherein the performance index is obtained after performance evaluation of the sub-model, and the consistency is determined based on the difference between the performance indexes of different sub-models.

14. A training device for a battery thermal runaway prediction model, characterized in that: include: A second data acquisition module is used to acquire training battery data and corresponding label information, and perform statistical analysis and feature construction on the training battery data to obtain statistical features and construction features accordingly; A training module is used to screen out target features related to thermal runaway from the statistical features and the structural features, and train each sub-model based on the target features and the label information to obtain a trained sub-model, wherein the sub-model is a model based on different machine learning algorithms, and the trained battery thermal runaway prediction model includes multiple trained sub-models.

15. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 8 or claims 9 to 12 are implemented.

16. A storage medium, characterized in that: The storage medium stores a plurality of instructions, which are suitable for being loaded by a processor to execute the steps of the method according to any one of claims 1 to 8 or claims 9 to 12.

17. A computer program product, characterized in that The invention comprises a computer program which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8 or claims 9 to 12.

Images