A novel battery abnormality detection method using interpretable Autoencoder. (15th January 2023)

• Record Type: Journal Article
• Authors: Zhang, Xiang; Liu, Peng; Lin, Ni; Zhang, Zhaosheng; Wang, Zhenpo
• Is Part Of: Applied energy. Volume 330:Part B(2023)
• Publication Date: 2023-01-15
• Subjects: Electric vehicles -- Lithium-ion battery pack -- Battery abnormal identification -- Autoencoder; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.120312
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: The proposed method is based on unsupervised learning, avoiding the problems of lack of samples; A knowledge-based guidance matrix is established to accelerate training speed and to ensure convergence; An interpretable Autoencoder is used to ensure extensibility among different vehicles; Effectiveness has been verified using data from real-world EVs. Abstract: The abnormality detection of lithium-ion battery pack is crucial to ensure the safety of electric vehicles (EVs). However, the dynamic and complex operating conditions of EVs making it challenging for algorithms designed under laboratory conditions to perform properly. In this study, a novel data-driven framework for abnormality detection is developed through establishment of a neural network with interpretable modules on top of an Autoencoder using data from real EVs to recognize abnormality while charging. The encoding guide matrix proposed in this method greatly accelerates the training speed, which also helps retains the learning ability of the neural network with consideration of the influence from each feature to provide supplementary information. The proposed algorithm is validated with data from real EVs. The results show that, compared with most existing algorithms, evidently higher accuracy can be achieved with shorter training time and lower computational cost, where the accuracy remains above 94% for all tested sample and the average root mean square error (RMSE) is as small as 0.03913. The proposed method can be utilized for both cloud-based and vehicle-based battery fault diagnoses.