Dictionary Learning Augmented Beamforming for Industrial Machine Inspection With Microphone Array

Acoustic signals are considered as one of the vital and early indicators of machine health. However, an observed acoustic signal acquired in industrial setting is highly corrupted by the interference and background noise. To address this problem, in this paper, we present a novel two-stage technique for acoustic-based machine anomaly detection. In the first stage, beamforming is employed for source separation at a coarser level. Subsequently, pre-trained dictionaries are used to estimate the individual source signals from the mixed signal. Once the sources are separated, a simple template matching approach is used to detect machine anomalies in the second stage. Performance evaluation is done using publicly available MIMII dataset that contains the machine sounds from different industrial machines. The results clearly indicate the efficacy of the proposed two-stage method for machine anomaly detection, compared to other signal processing and deep learning techniques.