Comparative Study of Excitation Signals for Active Fault Diagnosis of Belt Drives

Toothed belts are a popular drive solution in many industrial sectors. However, they are little noticed as fault diagnosis (FD) applications. For optimal operating conditions, the belt must be properly tensioned. An active FD combined with machine learning is pursued in this work to provide reliable belt looseness recognition. Active fault diagnosis increases the performance by injecting an additional excitation signal. This work addresses the two crucial steps of active, data-driven fault diagnosis which are (1) input signal design and (2) algorithm selection. For this purpose, test signals are investigated for an appropriate excitation. Based on the obtained data, FD is done by learning a time series regression (TSR) model. State-of-the-art TSR algorithms are benchmarked on multiple industrial datasets which are created by attaching different loads to the belt drive. In this way, we figure out how excitation and algorithm selection help to establish a safe and robust fault diagnosis that meets industrial requirements.