GRU-Attention Denoising Autoencoder Aided Fault Prognosis Method for System-level Application

Rolling bearings are important components in transmission systems, and the fault prognosis for them is crucial for the maintenance transition from prevention to prediction. However, previous studies based on component-level test data cannot be utilized in system-level applications since the precise incremental information of the signal is overwhelmed by the significant noise. To address these challenges, a fault prognosis method is proposed, aided by a denoising autoencoder that integrates a gated recurrent unit (GRU) model and multi-head attention. By introducing noise generated through dynamic simulation into the component-level test vibration signals, the network is enabled to segregate the relevant segments for prognosis from the health indicators (HI) of the noisy signals. Based on the envelope difference of the HI, the degradation path is categorized into healthy, slow degradation and rapid degradation stage. A prognosis model is then established to estimate the HI and current degradation stage of bearings using an evaluation indicator. With the proposed method, alarms can be triggered when degradation surpassing the permissible threshold set according to specific requirements in practical applications.