Electric vehicle multi-source fault diagnosis and early warning method based on decision tree model

Abstract

The invention discloses an electric vehicle multi-source fault diagnosis and early warning method based on a decision tree model, and belongs to the field of intelligent diagnosis and operation and maintenance of new energy vehicles. According to the method, various data sources generated in the running process of the electric vehicle are comprehensively utilized, and the data sources comprise battery management system (BMS) alarm data, running state label-free data and battery monomer cycle life data. Through feature engineering construction and mutual information analysis, feature variables strongly related to faults are extracted, an up-down sampling strategy is adopted to optimize data balance, and decision tree classification models are constructed for various typical faults respectively. After the model is trained, the model is deployed to a cloud platform, and online state monitoring, real-time fault early warning and alarm reason tracing are realized. The method has the advantages of being high in interpretability, high in diagnosis accuracy, convenient to deploy and the like, is suitable for an electric vehicle full-life-cycle health management system, and can remarkably improve operation safety and maintenance efficiency.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance of new energy vehicles, and in particular relates to a multi-source information fusion method that combines machine learning and data-driven approaches for the diagnosis and early warning of typical faults in electric vehicles. Background Technology

[0002] With the rapid development of the global new energy vehicle industry, electric vehicles, as clean energy transportation tools, have been widely used in urban travel, logistics and distribution, public transportation and other fields. Electric vehicles are characterized by high system complexity, variable operating conditions and diverse working environments, which leads to an increasingly diverse range of failures during actual use, especially in key components such as the power battery management system (BMS), motor controller, and on-board communication module, which are high-risk areas for failure.

[0003] In current mainstream electric vehicle operation and maintenance practices, fault diagnosis relies on manual inspection or static code analysis methods, which mainly have the following technical bottlenecks.

[0004] Low diagnostic efficiency and poor timeliness: Manual inspection usually intervenes after a fault occurs, making it difficult to respond to and predict faults in a timely manner, thus delaying the opportunity to deal with the problem.

[0005] Strong reliance on expert experience: Traditional rule-based methods rely excessively on domain experts' understanding of electrical structures, lacking uniformity and generalization ability.

[0006] Low data utilization: A large amount of structured and unstructured status data is generated during vehicle operation, but no effective integration and analysis mechanism has been formed, especially a large amount of unlabeled data is lost.

[0007] The problem of model "black box" is serious: Although algorithms such as neural networks have certain predictive capabilities, their results are difficult to interpret and difficult to assign responsibility for, which restricts their credibility in the engineering diagnosis process.

[0008] High deployment costs and poor portability: Some highly complex models are difficult to adapt to vehicle-mounted embedded systems or real-time cloud deployment, limiting their practical application scope.

[0009] The industry urgently needs an intelligent fault diagnosis method that combines high accuracy, low computational complexity, strong deployability, and good interpretability, which can integrate multi-source data from electric vehicles and adapt to the needs of industrial-grade diagnostics and cloud platform applications.

[0010] Against this backdrop, intelligent diagnostic methods based on decision tree models have become an important solution for remote fault diagnosis and health management systems for electric vehicles due to their clear structure, well-defined rules, and fast execution speed. By combining feature engineering, sample balancing, and model training strategies, decision tree models can effectively characterize the causal relationship between faults and multiple variables, and can visualize the structure within a controllable range, meeting the "transparent prediction" requirement in practical engineering applications.

[0011] Therefore, constructing a decision tree-based fault diagnosis and early warning method that integrates multi-source state data and possesses strong interpretability and real-time response capabilities is a key technical means to improve the safe operation capability of electric vehicles, extend their life cycle, and promote the development of intelligent connected vehicles. Summary of the Invention

[0012] This invention addresses the technical challenges of existing electric vehicle fault diagnosis methods, such as poor diagnostic timeliness, low level of intelligence, weak predictive ability, and uninterpretable black-box models. It proposes an intelligent fault diagnosis and early warning method based on a decision tree model that integrates multi-source operational data and possesses high interpretability and deployability. The aim is to achieve rapid location, effective early warning, and traceable cause analysis of typical electric vehicle faults, thereby improving the overall vehicle's intelligent operation and maintenance level and operational safety.

[0013] The objective of this invention can be achieved through the following technical solutions.

[0014] The multi-source data fusion method described above constructs a multi-channel data acquisition system for electric vehicle systems, integrating multi-dimensional time-series information such as BMS alarm data, controller operating status, historical untagged data, and battery cell charge-discharge cycle life data, providing a rich input foundation for diagnostic models.

[0015] The aforementioned intelligent preprocessing, based on data engineering methods, cleans, removes outliers, and normalizes the raw data, and uses the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to enhance fault category samples.

[0016] The aforementioned feature construction, combining domain knowledge and statistical methods, constructs and filters original and derived features. This mainly includes: constructing derived features (such as temperature rise rate, voltage change rate, SOC slope, etc.) through difference method, sliding statistics method and physical meaning modeling, and filtering features by combining mutual information analysis, feature importance ranking and other methods.

[0017] The training and evaluation of the decision tree model adopts the CART decision tree algorithm. A corresponding binary classification model is constructed for each type of typical fault. The model has a clear branch structure and interpretable decision boundaries. The generalization ability and stability of the model are evaluated through cross-validation.

[0018] The cloud platform deployment described above will deploy the trained decision tree model to the vehicle-cloud remote fault diagnosis platform, enabling online access and real-time inference of vehicle-side data.

[0019] The aforementioned early warning mechanism is constructed such that the diagnostic system triggers early warning pushes based on predicted probability thresholds, and the output results include fault type, related characteristics, possible causes and handling suggestions, and supports data visualization and regular incremental updates of the model.

[0020] The aforementioned visualization and feedback mechanism is implemented through a cloud platform integrated visualization module, which provides the following functions.

[0021] Fault trend monitoring charts, heat maps, and feature contribution values ​​are visually displayed.

[0022] A user feedback annotation mechanism is used to support subsequent online model updates.

[0023] Fault records are traceable and exportable, supporting operational and maintenance decisions.

[0024] Compared with the prior art, the present invention has the following beneficial effects.

[0025] I. Multi-source fault data fusion modeling mechanism: For the first time, data with different structures and time series during the operation of electric vehicles are modeled in a unified manner, which improves the information richness of the diagnostic model, has high fault identification accuracy, and can be adapted to a variety of typical fault types.

[0026] II. Highly Interpretive Decision Logic Design: Leveraging the advantages of decision tree structure, the diagnostic logic for each fault category is made transparent and traceable, overcoming the engineering adaptation challenges of neural network-type "black box" models; the decision tree model has the advantage of low computational complexity, which can effectively prevent data overfitting during the classification process.

[0027] III. Model Deployment and Early Warning Response Integration: Achieve seamless integration between the model and the actual remote operation and maintenance platform, with functions such as real-time data access, rapid response, remote alarm push and visual feedback.

[0028] IV. Extensible learning mechanism: Supports periodic retraining of the model and dynamic updating of the feature system to adapt to the evolutionary needs of different vehicle models, working conditions and environments. Attached Figure Description

[0029] To clearly illustrate the technical solution of the present invention, the accompanying drawings required in the description of the embodiments will be introduced below.

[0030] Figure 1 This is a flowchart illustrating the multi-source fault diagnosis and early warning method for electric vehicles based on the decision tree model of this invention. Detailed Implementation Plan

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0032] The present invention will now be described in further detail with reference to the accompanying drawings.

[0033] like Figure 1 As shown, the system first collects multi-source data during the operation of the electric vehicle through the on-board controller and sensor modules, including but not limited to: BMS alarm data, untagged operating status data, battery cell cycle life data, voltage, current, temperature, SOC, and historical vehicle operating trajectory data. This data is transmitted to the cloud platform via the on-board communication module, forming the basic information source for diagnostics.

[0034] The collected data is sent to the data cleaning and standardization module, where missing and outlier values ​​are removed and then normalized. To address the issue of uneven distribution of fault samples, the system employs the SMOTE algorithm to upsample minority fault samples, ensuring the balance and effectiveness of the training data.

[0035] For each sample in the minority class sample set From its k nearest neighbor samples One sample is randomly selected from the middle. , synthesize new samples.

[0036]

[0037] in, For minority class original samples, for The nearest neighbor samples; Randomly generated weighting coefficients are used in... and New samples are generated in the feature space.

[0038] Simultaneously, the Tomek Links method is used to undersample negative samples to optimize sample balance. Two samples (Majority class) and (Minority class) satisfies.

[0039]

[0040] but To form Tomek Links pairs, majority class samples need to be removed. .

[0041] in, For the sample and Euclidean distance.

[0042] Based on the preprocessed data, the system uses the feature engineering module to construct key variables, such as temperature rise rate, voltage change rate, SOC slope, maximum voltage difference, and current fluctuation frequency. Then, the features are screened by comparing mutual information and information gain, and several high-value variables that are most sensitive to faults are retained for the model to use.

[0043] feature With fault labels mutual information Defined as.

[0044]

[0045] in, Features The marginal probability distribution; Fault Label The marginal probability distribution; for and The joint probability distribution.

[0046] The system uses a decision tree modeling module, calling the CART algorithm to construct multiple binary classification diagnostic models for different fault types. The models employ the Gini index as the splitting criterion, learn fault discrimination paths through the training set, and undergo cross-validation to ensure their generalization ability. The model output structure is clear, with well-defined nodes, and possesses interpretability and engineering controllability.

[0047] The Gini index of the sample set (D).

[0048]

[0049] in, The total number of sample categories; for The Middle A subset of class samples Its sample size, for The total number of samples.

[0050] The Gini index of the subsets after splitting by feature A: if feature A will Divided into and Then the Gini index after splitting is .

[0051]

[0052] Choose to The smallest feature A is used as the splitting feature of the current node.

[0053] The resulting model is deployed on a cloud platform, enabling real-time data communication with the vehicle. When the system receives new status data, it triggers the model inference process to quickly generate diagnostic results, including predicted labels, confidence probabilities, and feature contributions, to determine whether the warning conditions have been met.

[0054] If the predicted probability exceeds the system's set threshold, the system immediately activates the early warning module, pushing structured alarm information to the driver's terminal and the operation and maintenance platform backend. The early warning content includes the fault type, associated characteristics, possible causes, and suggested remedial measures. The platform interface can display fault trend graphs and feature heatmaps in real time to assist technical personnel in responding quickly.

[0055] The system supports post-fault verification and annotation mechanisms, which can send the confirmation results from technical personnel back to the platform to build a retraining dataset, enabling online incremental learning and periodic updates of the model, and improving long-term accuracy and adaptability.

Claims

1. A method for multi-source fault diagnosis and early warning of electric vehicles based on a decision tree model, characterized in that, Includes the following steps: (1) Collect multi-source data during the operation of electric vehicles, including BMS alarm data, untagged operation data, battery cell cycle life data, etc. (2) Perform data preprocessing on the collected raw data, including upsampling, normalization, missing value imputation and outlier removal operations; (3) Based on the preprocessed data, construct derived features and extract highly correlated features through mutual information analysis, variance filtering and correlation screening methods; (4) Construct and train a binary classification decision tree model for each type of typical fault, and evaluate its performance metrics, including accuracy, recall and F1 score. (5) Deploy the trained model to a remote cloud platform to realize electric vehicle fault identification, early warning output and feedback mechanism, and support data visualization and online updates.

2. The method according to claim 1, characterized in that, The multi-source data in step (1) further includes operating condition data such as motor controller status, current fluctuation frequency, voltage ramp-up rate, and temperature gradient change.

3. The method according to claim 1, characterized in that, The upsampling and downsampling include using the SMOTE algorithm for positive sample oversampling and combining it with the Tomek Links method for negative sample undersampling to optimize sample balance.

4. The method according to claim 1, characterized in that, The mutual information analysis is used to measure the information gain between each feature variable and the fault label, and only the top N feature variables are retained by setting a threshold.

5. The method according to claim 1, characterized in that, The decision tree model is constructed using the CART (Classification and Regression Tree) algorithm, and the Gini index is used as the node splitting criterion.

6. The method according to claim 1, characterized in that, The model deployment module is integrated into the cloud-based fault diagnosis platform and achieves real-time data transmission with the vehicle terminal through the vehicle-cloud interface protocol.

7. The method according to claim 1, characterized in that, The warning output includes the fault type, trigger time, suspected component, predicted probability, and handling suggestions, and the information is pushed through mobile terminals.

8. The method according to claim 1, characterized in that, After deployment, the model supports an incremental learning and update mechanism, which performs periodic retraining based on newly collected labeled data to improve the model's diagnostic accuracy.

9. The method according to claim 1, characterized in that, The visual feedback interface uses charts to display the importance indicators of each key feature, the current vehicle status, and the historical fault distribution trend.

10. The method according to claim 1, characterized in that, This method is applicable to fault diagnosis platforms for different vehicle models. By configuring specific preprocessing parameters and model tuning interfaces, cross-platform migration and deployment can be completed.

Images