A speed normalized autoencoder for rotating machinery fault detection under varying speed conditions

Abstract

Rotating machinery often operates under varying speed conditions. Fault detection is necessary to prevent sudden failures and enable condition-based maintenance. Existing autoencoder-based (AE-based) fault detection methods did not address the effects of speed variations, and thus leave room for improvement at varying speed conditions. This paper proposes a new deep learning model named speed normalized autoencoder (SN-AE). The SN-AE consists of a speed normalization (SN) branch and an AE branch. The SN branch takes the speed signal as the input and automatically learns an SN function which normalizes the vibration signal to remove the effects of speed variations. Thereafter, the normalized vibration signal is inputted to the AE branch for fault detection. Case studies were conducted to detect incipient faults of three typical rotating machines including a planetary gearbox, a fixed-shaft gearbox and a rolling element bearing under varying speed conditions. Results have shown that the proposed SN-AE successfully removes the effects of speed variations and achieves significantly better detection performances than existing AE-based fault detection methods.