A Data-Selection Framework for Data-Efficient Battery Parameter Estimation

Online battery parameter identification is critical for accurate monitoring of battery states. However, conventional identification methods will perform poorly when the measured cell data is not informative enough. Data selection methods can be used to detect and only use high-quality cell data segments, thus making the identification process more efficient. In this paper, we measure the quality of a data segment based on the Cramer-Rao lower bound under the Total Least Squares (TLS) framework. A Convolutional Neural Network (CNN) is then used to predict if a data segment is useful or not based on the acceleration profile of the vehicle. The CNN was trained and validated on synthetic cell data generated from simulated trips in the city of Berlin. The results show good performance of the CNN and the proposed data selection algorithm yields cell models which performs better than conventional identification methods with a data reduction of up to 56 percent.