Unsupervised fault detection using frequency-wise angular filtering in contaminated vibration signals

Manufacturing processes involve multiple machines within a production line. Unexpected faults in machines reduce productivity and increase maintenance costs. Engineers face difficulties in managing numerous machines individually and controlling them immediately. For automatic condition monitoring, several studies have focused on multivariate statistical process control and fault detection based on artificial intelligence. These methods require labeled data or assume that the training data contains only normal patterns. However, obtaining labeled data in the industry is challenging because engineers must manually label the data. Contaminated signals containing fault patterns in unlabeled training data significantly degrade the performance of fault detection in the model. This study proposes Unsupervised fault detection with Frequency-wise Angular Filtering (UFAF) to improve the performance of fault detection in contaminated vibration signals. The UFAF extracts angular features to estimate the normal samples for use only during model training. This filtering strategy is repeated at every epoch and is eventually optimised to use only high-quality normal samples during model training. An experiment using SpectraQuest gearbox datasets confirms the excellent performance for contaminated signals, as angular features are effective in identifying normal and fault signals. The UFAF is practical and applicable in industries wherein it is difficult to collect labeled data.