Risk Assessment of Vehicle Battery Safety based on Abnormal Features and Neural Networks

Abstract

In this study, we evaluate a proactive battery EV safety assessment method using abnormal feature detection and neural networks. Four sophisticated algorithms —Isolation Timberland, One-Class SVM, Autoencoder and also LSTM— were performed to assess their applicability in detecting anomalous battery behavior. The Isolation Woodland algorithm showed a balanced accuracy recall trade-off of the values 0.85 and 0.92 respectively One class SVM demonstrated highly sharp results with an accuracy and recall values of 0.78 and 0.8, respectively. The autoencoder, that used a large amount of learning and won with 0.92 accuracy score and an F1-score – 0.89 The LSTM structure, programmed for sequential information, indicated a great execution with a 0.94 review and the F1-score of 0. A comparative study has shown that these algorithms can provide alot flexibility in sending based on the clear requirements.