Abrupt variance and discernibility analyses of multi-sensor signals for fault pattern extraction

High performance sensors and data logging technologies have enabled us to monitor the operational states of systems and predict fault occurrences. Because a large number of sensors are monitored, it is important to identify significant sensor signals for detecting faults. As signal behaviors become increasingly scattered and complex, it is difficult to investigate the gradient relationship with the operational states of a system; therefore, we analyze the abrupt variance and the discernibility index of multi-sensor signals by extending the conventional statistical variance and the Fisher criterion. Based on the two novel characteristics of sensor signals, we select the most significant sensors to detect abnormal cylinder temperature and engine knocking. Thus, the proposed analyses lead to improved detection results when the multi-sensor signals existed in multiple overlapping regions regardless of the operational states.