Precision-Concentrated Battery Defect Detection Method in Real-World Electric Vehicles Crossing Different Temperatures and Vehicle States

Hundreds of electric vehicle (EV) battery thermal runaway accidents resulting from untreated defects restrict further development of EV industry. Battery defect detection based on the abnormality of external parameters is a promising way to reduce this kind of thermal runaway accidents and protect EV consumers from fire danger. However, the influence of temperature and EV states, i.e., charging and driving, on the battery characteristic will complicate the method establishment. Existing data-driven methods are continuously falsely judging normal batteries to be defected, which will cause panic of EV occupants. To cope with the issue, a precision-concentrated battery defect detection method crossing different temperatures and vehicle states is constructed. The method only uses sparse and noisy voltage from existing onboard sensors. First, a density-based semi-supervised cluster (DBSSC) method is proposed containing three novelties: the objective function is originally defined and a multilayer L-shaped optimization method (MLOM) is proposed to improve the precision; six kernel domains are proposed to cope with the arbitrary distribution of battery voltages; the soft boundary is designed to consider the random noise in real-world EV operation. Subsequently, the DBSSC is trained by real-world data of different EV states and temperatures to enhance the robustness. The training process only uses data of normal batteries to cope with the inadequacy of thermal runaway battery data. The results show that the method can detect defected batteries 13 days ahead the thermal runaway while achieve the precision of 99.2%. By the three novelties and training by data of different conditions, the precisions are improved by 40.9%, 3.4%, 7.0%, and 12.0%, respectively.