An Analytics Methodology for Anomaly Detection In Machine Log Data

Throughout the production of wafer in semiconductor manufacturing, a large number of log files are created. The information in logs generated by devices and applications can be used to assess the system’s health. The analysis of logs is crucial to verify whether equipment went smoothly. Manual review is insufficiently trustworthy, and takes much time and effort. In this paper, we present an algorithm for anomaly detection and raise alarms for potential incidents on time. In actual manufacturing, there are various abnormal conditions, including single point mutation, long-term trend change, anomalous divergence and abrupt change in numeric level, which are short of manual labeling. As a result, we innovatively modify and ensemble multiple segmentation algorithms, such as Mann-Kendall (M-K), Bernaola Galvan (BG) and Pettitt mutation test, and applied unsupervised learning i.e. Spectral Residual (SR) to better improve the performance. The results achieved by our approach are superior to other conventional time series anomaly detection methods.