Anomaly Detection in Battery Charge-Discharge Experiments Using Multi-Task Auto-Encoder

Abstract

Listen

Translate article

Battery charge-discharge experiments require a long time for data collection, often taking several months. It is impossible to stop the experiment immediately when anomalous data is generated, such as rapid capacity fade in batteries, leading to inefficient waste of time.To effectively continue the experiment, it's important to monitor the battery charge and discharge process in real time to anomaly behavior early.Recently, research has been actively conducted to predict the characteristics and lifetime of batteries using deep learning. However, this deep learning approach necessarily requires a large amount of reproducible learning data. The fact that battery data is highly affected by changes in experimental conditions such as C-rate and temperature.In this paper, we propose an anomaly detection algorithm model applicable to various experimental conditions by using multiple Auto-encoders (AE). We define this model as Multi-task Auto-encoder (MTAE), where each AE model is trained on normal battery operation data collected under different C-rate conditions.MTAE reconstructs raw data measured at each cycle during charge-discharge experiments based on the learned normal data and calculates the reconstruction error. Using the reconstruction error values, we establish the 'reconstruction error distribution' for normal battery operation data, and based on this, set a threshold for detecting anomaly data. If the reconstruction error exceeds the threshold, it's identified as anomaly data.Real-time anomaly detection using raw data through MTAE showed higher performance compared to AE models trained solely on a single C-rate, as evaluated by the Area Under the Curve (AUC). Although this paper categorized data based only on C-rate, future research will extend to include various characteristics such as temperature and pressure.