Battery fault prediction method under new energy electric vehicle scene

Abstract

The application relates to the technical field of big data, in particular to a battery fault prediction method in a new energy electric vehicle scene, which comprises the following steps: collecting electric vehicle battery related information data sets and preprocessing the same to obtain preprocessed data sets; based on the preprocessed data sets, an extended TrAdaBoost algorithm is used to construct a classifier; knowledge transfer is realized by using the classifier to obtain new data sets; feature engineering is performed on the new data sets, including basic features and business features; a replacement feature importance method is used to screen features; the screened features are input into an AutoGluon framework, and a battery fault prediction result is output; the extended TrAdaBoost algorithm is used to combine data in two different fields, and the knowledge of a source field is transferred to a target field through the idea of transfer learning, so that the generalization performance of a model in the target field is enhanced.

Description

Technical Field

[0001] This invention relates to the field of big data technology, specifically to a method for predicting battery failures in the context of new energy electric vehicles. Background Technology

[0002] Driven by increasingly stringent emission standards and ever-improving environmental regulations, electric vehicles are being vigorously promoted and popularized as an effective way to reduce air pollution and carbon emissions. The electric vehicle battery is the core component of an electric vehicle. With advancements in lithium-ion battery technology, the energy density, charging speed, and lifespan of electric vehicle batteries have significantly improved. Simultaneously, the development of solid-state batteries, fast-charging technology, and other new battery technologies is continuously driving the development of electric vehicles. However, this also brings the issue of electric vehicle battery safety. Battery failures often lead to vehicle malfunctions, explosions, and other safety accidents. Therefore, predictive management of electric vehicle battery failures is essential.

[0003] Battery failure prediction aims to improve the reliability and safety of electric vehicles by monitoring battery operating status and environmental conditions to provide early warnings of potential battery failures. Currently, the main methods for battery failure prediction include data-driven machine learning algorithms and physical modeling methods. Data-driven methods primarily utilize historical data (such as voltage, temperature, and state of charge) to train predictive models, which typically include regression models, classification models, and time-series models. Physical modeling methods are based on the physical and chemical characteristics of the battery, establishing mathematical models to simulate battery behavior. These models can provide a deep understanding of battery failure mechanisms, but usually require detailed battery parameters and high computational complexity. Furthermore, current battery failure prediction methods often only use data from electric vehicles of the same brand for model training and prediction. While these models perform well in relatively fixed environments, they often perform poorly under sudden conditions. Specifically, current battery failure prediction faces the following challenges:

[0004] 1) Limited data: In the electric vehicle market, fault data for specific brands of electric vehicles may be relatively limited, especially for newly launched brands or models. Data scarcity leads to insufficient training data for the model, thus affecting predictive performance;

[0005] 2) Imbalanced sample size: Battery failure events are usually rare in the data, which may result in insufficient failure samples in the training data, affecting the learning effect of the model.

[0006] 3) Data homogeneity: Battery design, manufacturing processes, and operating conditions may differ between different brands of electric vehicles. Using data from a single brand for training may result in the model failing to generalize to batteries from other brands.

[0007] 4) Poor model generalization ability: Traditional models often need to be retrained or adjusted when faced with data from different brands or types of batteries, which increases the complexity of model deployment. Summary of the Invention

[0008] To address the above problems, this invention provides a battery fault prediction method for new energy electric vehicles, characterized by the following steps:

[0009] S1. Collect a dataset of information related to electric vehicle batteries and preprocess it to obtain a preprocessed dataset. The dataset of information related to electric vehicle batteries includes battery data of the target electric vehicle brand and battery data of other electric vehicle brands.

[0010] S2. Based on the preprocessed dataset, a classifier is constructed using the extended TrAdaBoost algorithm; the classifier is used to filter the preprocessed battery data of other electric vehicle brands to obtain filtered data, and the filtered data is combined with the preprocessed battery data of the target electric vehicle brand to obtain a new dataset;

[0011] S3. For each new data point in the new dataset, construct 5 features, including basic battery features, battery business features, time features, discrete features, and cross features;

[0012] S4. Use the permutation feature importance method to select features from all features corresponding to the new dataset;

[0013] S5. Input the selected features into the AutoGluon framework and output the battery failure prediction results.

[0014] The beneficial effects of this invention are:

[0015] This invention employs an extended TrAdaBoost algorithm, combining data from two different domains. Through transfer learning, it transfers knowledge from the source domain (battery data from other electric vehicle brands) to the target domain (battery data from the target electric vehicle brand), enhancing the model's generalization performance in the target domain. When combining battery data from different electric vehicle brands, to address the issues of imbalanced samples and unequal contributions from the source and target domains, the definition of weights is adjusted during weight initialization. Furthermore, feature weights are introduced to adjust the contributions of different features when processing the relationship between features in the source and target domains.

[0016] This invention employs a method combining the importance and stability selection of permutation features to screen features. It also extends the original method by considering the synergistic effects between features and addressing the relationships between them. Multiple random samplings are used to examine the consistency and stability of features, reducing the risk of overusing the importance of permutation features, improving the reliability of the screening results, and retaining as few important features as possible, thus improving efficiency while saving costs. Attached Figure Description

[0017] Figure 1 This is a flowchart of a battery fault prediction method in a new energy electric vehicle scenario according to an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the TrAdaBoost algorithm flow for a battery fault prediction method in a new energy electric vehicle scenario according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a battery fault prediction method in the context of new energy electric vehicles, such as... Figure 1 As shown, it includes the following steps:

[0021] S1. Collect a dataset of information related to electric vehicle batteries and preprocess it to obtain a preprocessed dataset. The dataset of information related to electric vehicle batteries includes battery data of the target electric vehicle brand and battery data of other electric vehicle brands.

[0022] Specifically, the battery data for the target electric vehicle brand includes historical usage data for multiple batteries under the target battery brand; the battery data for other electric vehicle brands includes historical usage data for multiple batteries under other battery brands; the historical usage data includes battery life, battery failure time, battery failure cause, number of battery failures, etc.

[0023] Specifically, to improve data quality and applicability and reduce the risk of overfitting, preprocessing of the electric vehicle battery-related information dataset is necessary. Preprocessing any single data point in the electric vehicle battery-related information dataset includes:

[0024] S11. Clean the data; if the data has missing values, proceed to step S12; if the data has outliers, proceed to step S13; if the data has neither missing values ​​nor outliers, proceed to step S14.

[0025] S12. Use the KNN method to fill in the missing values ​​to form a new data line, and then proceed to step S14.

[0026] Specifically, in step S12, when using the KNN method to fill in missing values, weights are set to increase the proportion of values ​​above and below the missing position, as specifically expressed as follows:

[0027]

[0028] Among them, X missing Indicates missing values; X above This represents the value at the missing position, that is, the value at the position preceding the missing value; X below ω1 represents the value below the missing position, i.e., the value at the position immediately following the missing value; ω2 represents the weight of the value at the missing position, ω3 represents the weight of the value below the missing position, K represents the data length, and X represents the missing value. k This represents the value at the k-th position in the data. In this invention, ω1 = 0.4, ω2 = 0.4, and ω3 = 0.2.

[0029] S13. Draw a box plot based on all fields in the data. Calculate the interquartile range IQR = q3 - q1 based on the upper quartile q1 and lower quartile q3 in the box plot. Calculate the limit range [q1 - 1.5IQR, q3 + 1.5IQR] based on the interquartile range IQR. Delete fields outside the limit range to form a new data entry, and then proceed to step S14.

[0030] S14. Tag encoding is performed on the text content of this data, and the date content of this data is converted into a timestamp to extract the year and month, resulting in a preprocessed data.

[0031] S2. Based on the preprocessed dataset, a classifier is constructed using the extended TrAdaBoost algorithm. The classifier is used to filter the preprocessed battery data of other electric vehicle brands to obtain filtered data. The filtered data is then combined with the preprocessed battery data of the target electric vehicle brand to obtain a new dataset. The purpose of using the classifier for filtering is to determine whether the data of other electric vehicle brands can be added to the battery data of the target electric vehicle brand and used together as a training set to predict the test set. Relevant evaluation indicators are designed based on the battery data of the target electric vehicle brand to make the judgment.

[0032] Specifically, such as Figure 2 As shown, step S2, which uses the extended TrAdaBoost algorithm to construct a classifier, includes:

[0033] S21. Combine the preprocessed battery datasets of other trolley brands and the preprocessed target trolley brand battery dataset to construct the first dataset. All data in the first dataset containing the label "isDefault" form the training set T, and the remaining data form the test set S; where... X b This represents the preprocessed battery data for the target electric vehicle brand, X. a This represents preprocessed battery data from other electric vehicle brands, and Y represents the isDefault category space.

[0034] S22. Subdivide the training set into source domain datasets. and target domain dataset in Represents source domain data The true category, Represents target domain data The true category.

[0035] S23. Set initial weights for each data point in the training set to obtain the initial weight vector. Among them W i 1 represents the initial weight of the i-th data point, n represents the number of data points in the source domain dataset, and m represents the number of data points in the target domain dataset.

[0036] Specifically, considering the imbalance of samples and the significant difference in data volume between the source and target domains, different initial weights are set; the initial weights are calculated as follows:

[0037]

[0038] Among them, C a C represents the number of categories in the source domain dataset. b Indicates the number of categories in the target domain dataset, class weight Indicates the overall sample weight. This represents the category weight of the i-th source domain data item. This represents the category weight of the i-th target domain data.

[0039] S24. Perform multiple iterations of training on the initial weight vector to obtain the classifier.

[0040] Specifically, the initial weighting coefficient β0 is set as follows:

[0041]

[0042] Where N represents the number of iterations.

[0043] Specifically, each iteration of training includes:

[0044] S241. Update weight distribution p t The updated formula is:

[0045]

[0046] in, Let W represent the weight vector for the t-th iteration of training, where W... i t This represents the weight of the i-th data point during the t-th iteration of training;

[0047] S242. Call the basic algorithm Learner, based on the training set T and weight distribution p. t A classifier h is obtained from the test set S. t :

[0048] S243. Calculate the classifier h t In the target domain dataset T b Classification error on ∈ t :

[0049]

[0050] in, For the classifier h in the t-th iteration t For target domain data The prediction results For target domain data The true category;

[0051] According to the weight vector W t To adjust the contributions of different features and calculate the loss function L t :

[0052]

[0053] in, Representing data The loss function in the t-th iteration, Representing data The true label, The model represents the data. The predicted value is λ, which is a parameter used to balance the loss contributions of the source and target domains, and has a value of 0.4. By minimizing the loss function in each iteration, the feature weights and sample weights are gradually adjusted to reduce the overall classification error rate and lower the classification error value.

[0054] S244. Update weight coefficient β t The updated formula is:

[0055]

[0056] The updated weight vector for S245 is represented as follows:

[0057]

[0058] in This represents the weight of the i-th data item after the update.

[0059] Specifically, after N iterations, the final classifier is obtained by balancing the contributions of the source and target domains:

[0060]

[0061] S3. For each new data point in the new dataset, construct 5 features, including basic battery features, battery business features, time features, discrete features, and cross features;

[0062] Specifically, basic battery characteristics include fundamental information such as battery brand and capacity; battery business characteristics are obtained by processing the basic battery data, such as calculating resistance and whether the voltage exceeds a certain value over a continuous period; time characteristics include features such as sliding window features based on time. Discrete and cross-features are also battery-related data constructed and obtained using conventional methods.

[0063] S4. Use the permutation feature importance method to select features from all features corresponding to the new dataset.

[0064] Specifically, step S4 uses the permutation feature importance method to screen features, including:

[0065] S41. Calculate the stability score for each feature, expressed as:

[0066]

[0067] Among them, S j Let B represent the stability score of feature j, B represent the number of bootstraps, and I(·) represent the indicator function.

[0068] S42. Calculate the permutation feature importance score for each feature, expressed as:

[0069]

[0070] Among them, PI j The permutation feature importance score represents feature j, M represents the number of permutations, and Δimprtance m This represents the change in performance indicators after the m-th permutation.

[0071] S43. Calculate the overall score for each feature, expressed as:

[0072] score j =ω·S j +(1-ω)PI j

[0073] Among them, score j ω represents the overall score of feature j, and ω represents the weighting parameter for balancing stability and the importance of the permutation feature;

[0074] S44. Each feature forms a scoring relationship pair based on its own name and overall score, represented as follows:

[0075] Num j =(name j score j )

[0076] Among them, Num j Represents the score relation pair for feature j, name j Indicates the name of feature j;

[0077] S45. All features are sorted in descending order of comprehensive score, the first 4 / 3 of the features are retained, and the last 4 / 1 of the features are removed to reduce feature dimensionality and save computational resources.

[0078] S5. Input the selected features into the AutoGluon framework and output the battery failure prediction results.

[0079] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A battery fault prediction method in the context of new energy electric vehicles, characterized in that, Includes the following steps: S1. Collect a dataset of information related to electric vehicle batteries and preprocess it to obtain a preprocessed dataset. The dataset of information related to electric vehicle batteries includes battery data of the target electric vehicle brand and battery data of other electric vehicle brands. S2. Based on the preprocessed dataset, a classifier is constructed using the extended TrAdaBoost algorithm; A classifier is used to filter the preprocessed battery data of other electric vehicle brands to obtain filtered data. The filtered data is then combined with the preprocessed battery data of the target electric vehicle brand to obtain a new dataset. Step S2, which uses the extended TrAdaBoost algorithm to construct a classifier, includes: S21. Combine the preprocessed battery data of other EV brands with the preprocessed battery data of the target EV brand to construct the first dataset. All data in the first dataset containing the label "isDefault" will form the training set T, and the remaining data will form the test set S; where T... {X=X b ∪X a ×Y}, X b This represents the preprocessed battery data for the target electric vehicle brand, X. a This represents preprocessed battery data from other electric vehicle brands, where Y represents the isDefault category space; S22. Subdivide the training set into source domain datasets. and target domain dataset ;in Represents source domain data The true category, Represents target domain data The true category; S23. Set initial weights for each data point in the training set to obtain the initial weight vector. ,in The initial weight of the i-th data point is represented by n, the number of data points in the source domain dataset is represented by n, and the number of data points in the target domain dataset is represented by m. S24. Perform multiple iterations of training on the initial weight vector to obtain the classifier; S3. For each new data point in the new dataset, construct 5 features, including basic battery features, battery business features, time features, discrete features, and cross features; S4. Use the permutation feature importance method to select features from all features corresponding to the new dataset; S5. Input the selected features into the AutoGluon framework and output the battery failure prediction results.

2. The battery fault prediction method in the scenario of new energy electric vehicles according to claim 1, characterized in that, Preprocessing any single data point in the electric vehicle battery-related information dataset includes: S11. Clean the data; if the data contains missing values, proceed to step S12; if the data contains outliers, proceed to step S13; if the data contains neither missing nor outliers, proceed to step S14. S12. Use the KNN method to fill in the missing values ​​to form a new data line, and then proceed to step S14; S13. Draw a box plot based on all fields in the data. Calculate the interquartile range IQR = q3 - q1 based on the upper quartile q1 and lower quartile q3 in the box plot. Calculate the limit range [q1 - 1.5IQR, q3 + 1.5IQR] based on the interquartile range IQR. Delete fields outside the limit range to form a new data entry, and then proceed to step S14. S14. Tag encoding is performed on the text content of the data, and the date content of the data is converted into a timestamp to obtain a preprocessed data.

3. The battery fault prediction method in the scenario of new energy electric vehicles according to claim 2, characterized in that, In step S12, when using the KNN method to fill in missing values, weights are set to increase the proportion of values ​​above and below the missing position, specifically as follows: Among them, X missing Indicates missing values; X above This represents the value at the missing position, that is, the value at the position preceding the missing value; X below ω1 represents the value below the missing position, i.e., the value at the position immediately following the missing value; ω2 represents the weight of the value at the missing position, ω3 represents the weight of the value below the missing position, K represents the data length, and X represents the missing value. k This represents the value at the k-th position in the data.

4. The battery fault prediction method in the scenario of new energy electric vehicles according to claim 1, characterized in that, The initial weight calculation formula is: Among them, C a C represents the number of categories in the source domain dataset. b Indicates the number of categories in the target domain dataset, class weight Indicates the overall sample weight. This represents the category weight of the i-th source domain data item. This represents the category weight of the i-th target domain data.

5. The battery fault prediction method in a new energy electric vehicle scenario according to claim 1, characterized in that, Each iteration of training includes: S241. Update weight distribution p t The updated formula is: in, Let represent the weight vector for the t-th iteration of training, where This represents the weight of the i-th data point during the t-th iteration of training; S242. Call the basic algorithm Learner, based on the training set T and weight distribution p. t A classifier h is obtained from the test set S. t :X Y; S243. Calculate the classifier h t In the target domain dataset T b Classification error : in, For the classifier in the t-th iteration For target domain data The prediction results For target domain data The true category; Based on the weight vector To adjust the contributions of different features and calculate the loss function. : in, Representing data The loss function in the t-th iteration, Representing data The true label, The model represents the data. The predicted value, This is a parameter used to balance the loss contributions of the source and target domains, and its value is 0.4; S244. Update weight coefficient β t The updated formula is: The updated weight vector for S245 is represented as follows: in This represents the weight of the i-th data item after the update.

6. The battery fault prediction method in a new energy electric vehicle scenario according to claim 1, characterized in that, Step S4 uses the permutation feature importance method to screen features, including: S41. Calculate the stability score for each feature, expressed as: Among them, S j Let B represent the stability score of feature j, B represent the number of bootstraps, and I(·) represent the indicator function. S42. Calculate the permutation feature importance score for each feature, expressed as: Among them, PI j The permutation score represents the feature importance score of feature j, and M represents the number of permutations. This represents the change in performance indicators after the m-th permutation. S43. Calculate the overall score for each feature, expressed as: Among them, score j ω represents the overall score of feature j, and ω represents the weighting parameter for balancing stability and the importance of the permutation feature; S44. Each feature forms a scoring relationship pair based on its own name and overall score, represented as follows: Among them, Num j Represents the score relation pair for feature j, name j Indicates the name of feature j; S45. All features are sorted in descending order of their overall scores, and the first 4 / 3 of the features are retained while the last 4 / 1 are removed.

Images