Anomaly detection of vectorized time series on aircraft battery data
Moting Su, Wenjie Zhao, Ye Zhu, Donglan Zha, Yushu Zhang, Peng Xu · Expert Systems with Applications · 2023
The power supply system, as an indispensable electronic hardware module in most vehicles, needs the highest level of security and reliability to ensure the normal operation of the vehicle. Efficiently identifying any faulty battery at the earliest stage would prevent potential safety hazards. This paper aims to detect anomalous batteries using a time series analysis of their internal resistance. To identify the most meaningful patterns and extract their features, we propose a method named Pattern-based Vectorization for Time series (PVT). The PVT first encodes the local sequential shapes as unique symbols by sliding window, then maps each time series into a sequence of representative symbols, and finally generates the symbolic feature matrix for the whole time-series data via a TF-IDF statistical method. The effectiveness of PVT has been systematically evaluated with 7 classifiers on a large real civil aviation battery dataset collected from an uninterruptible power system. Our results show that PVT can significantly improve the performance of existing classifiers in detecting abnormal batteries. In particular, the combination with the RUBT classification model achieves the highest performance due to the random undersampling and boosting techniques that suit imbalanced data in anomaly detection scenarios.