Multitimescale Feature Extraction from Multisensor Data Using Deep Neural Network for Battery State-of-Charge and State-of-Health Co-Estimation

Accurate state estimation is necessary for battery management systems (BMSs) in electric vehicles (EVs) to deploy appropriate control policy; thus, the safety of the battery pack can be ensured, and lifespan can be prolonged. Current state estimation methods cannot fully exploit the battery multisensor data from a multitimescale perspective, which results in deteriorating estimation performance in laboratory testing data, let alone real-world application scenarios. To overcome the above drawbacks, this article proposes a deep neural network-based state-of-charge (SoC) and state-of-health (SoH) co-estimation framework, which could realize accurate estimation in both laboratory and realistic scenes. To realize multisensor data fusion, the original data is rearranged into a two-dimensional (2-D) matrix, with one dimension representing the time domain and the other representing the feature domain. To exploit the multitimescale changing properties related to SoC and SoH, convolutional filters with different sizes are used to extract features in different timescales; furthermore, the swish activation function and long short-Term memory (LSTM) layer are introduced to enhance the network convergence and estimation accuracy. The global average pooling (GAP) layer is adopted to substitute the traditional fully connected (FC) layer for network lightweight. Oxford public battery dataset and real-world EV battery operational data are used to verify the applicability of the proposed method. Results show that the SoC and SoH estimation errors are 1.43% and 1.59%, respectively, for the Oxford dataset, which is superior to many existing advanced machine learning models. In addition, the SoC and pseudo-SoH estimation errors in real-world EV driving scenarios are 0.79% and 2.59%, respectively, further verifying the accuracy and generalization capability of the proposed method.