Early Fault Detection Using Deep Learning on Compressed Cyclic Spectral Coherence Maps

Research nowadays takes advantage of the cyclostationary properties in the vibration waveforms of rotating machines for fault detection. For example, cyclic spectral coherence maps (CSCM) break down vibration signals into cyclic and carrier frequencies. However, the large size of the CSCMs and the enormous amount of data makes it challenging to identify the faults. Therefore this paper presents an approach that uses compressed CSCMs as input for a deep learning autoencoder, as a means to detect mechanical failures from vibration signals. The autoencoder learns to reconstruct a healthy version of the CSCMs, taking advantage of the generalization bias due to the unbalanced existence of healthy data. Furthermore, if the autoencoder receives a faulty CSCM as input, it reconstructs a healthy version. Alarms are triggered if the residuals between reconstruction and input CSCM exceed a calculated threshold. Next, the reporting step lists the cyclic order, carrier frequency band, and their respective alarm numbers. Five years of data from five different turbines in the field is used to train the autoencoder. The evaluation is validated on data from two faulty and eight healthy turbines. This work provides a fast, reliable and automated method to detect mechanical failures.