| As an important power source of electric vehicles,lithium-ion battery system has a decisive influence on the safety and reliability of electric vehicle operation.The life prediction and fault diagnosis for lithium-ion batteries play a critical role.There are many cases in which the abnormality of lithium-ion batteries under various time scales does harm to the the safety and reliability of electric vehicle operation.The dissertation focuses on the abnormal aging identification of lithium-ion batteries and the voltage signal fault diagnosis of the battery system in electric vehicles.The research was conducted on the life prediction of lithium-ion batteries with accelerated capacity degradation using capacity data by aging test,and battery pack fault diagnosis using historical fault data in practical vehicle operation.The main objective of this dissertation is to improve the performance of existing battery life prediction and fault diagnosis methods.The content is as follows:Firstly,to achieve life prediction of lithium-ion batteries with accelerated capacity degradation,an empirical model for accelerated capacity degradation is proposed,which shows high data fitting accuracy and parameter identifiability;the system noise covariance matrix in the state space equation of the unscented particle filtering algorithm is modified to express the conversion of the dominant term of the model after the inflection point identification.The accelerated capacity degradation point can be predicted several to a dozen cycles in advance,and the prediction accuracy is improved by the proposed method,compared with the traditional particle filtering algorithm using a double exponential empirical model,which indicates the theoretical significance of the proposed method on the reliability of electric vehicle operation.Secondly,to further reduce the time-complexity and errors of the model-based method,the sources of model-based prediction errors are analyzed,and a hybrid method based on error correction is proposed for predicting the accelerated capacity degradation of lithium-ion batteries to further improve the performance of the prediction.Empirical mode decomposition is used to analyze the original error series,correlation analysis is performed on the IMFs to reconstruct the error series,then Gaussian process regression is used to predict subsequent error series,and the prediction results of the error series are used to correct the model-based prediction results.Using the same training set as in the model-based study,the prediction accuracy is further improved,with the vast majority of the verification dataset falling into the 95% confidence intervals of the prediction results,and the time consumption is reduced by the proposed method.Moreover,the practical availability of the hybrid method is discussed,and the necessity of the research on short time scale battery dataset is then illustrated.Thirdly,considering relatively low quality and resolution of the battery dataset provided by the big data platform,by using the characteristics of the charging segments of historical vehicle data which are relatively smooth,a battery cell inconsistency quantification method based on the similarity of the voltage time series of the charging segments is proposed,and the alignment process of the voltage time series is obviously optimized by using the DTW distance compared with Euclidean distance;meanwhile,a nonlinear mapping strategy is proposed to realize identification and early warning of battery pack inconsistency,based on the evolutionary trajectory of gradual degradation of the battery cell consistency by horizontal and vertical comparison.Finally,to improve the accuracy and sensibility of existing battery fault diagnosis methods,by adopting variational mode decomposition,the bias of the voltage signal in the discharge segments caused by state inconsistency of the pack at long time scales can be eliminated,and the performance of the steady-state component elimination is verified to become significantly enhanced,compared with adopting empirical mode decomposition;a generalized dimensionless indicator construction formula for the voltage signal is proposed,and it is demonstrated that the typical signal features in existing signal-based battery fault diagnosis research are proved to be particular cases of the proposed formula;based on the feature sequences constructed,a clustering-based anomaly detection method is proposed to detect voltage signal anomalies,which can identify the type,moment of occurrence,duration and anomaly severity of local anomalous signal patterns.The signal features used in existing fault diagnosis methods are difficult to achieve the performance at the same time.Combined with the Manifold learning,an anomaly detection method with feature dimensionality reduction is proposed for multiple voltage signal feature analysis under various fluctuating operating conditions,which could make less misjudgment compared with using PCA.From the perspective of short time scale,the proposed fault diagnosis method can achieve fault warning minutes to hours earlier than the fault diagnosis of the battery management system in the verified cases,reducing the false positive diagnosis results of traditional thresholds;from the perspective of long time scale,the utilization of time-weighted standardized feature sequences can effectively capture and accumulate the minor abnormalities that occur intermittently in the battery cells,achieving early abnormality detection,which indicates the theoretical significance of the method on the reliability of electric vehicle operation.To sum up,the study on the life prediction and fault diagnosis of lithium-ion batteries is conducted in the dissertation,and the existing relevant methods are summarized.Making full use of the battery dataset under various time scales,the proposed methods are proven to improve and enhance the performance of existing methods,thus the research contains theoretical value and practical significance on the safety and reliability of electric vehicle operation. |
PDF Full Text Request