Sound Database of Industrial Machine for Audio Anomaly Detection

Audio anomaly detection (AAD) can seamlessly determine faults in industrial machines and improve the efficiency of predictive maintenance systems. However, the unavailability of audio sound recordings of real industrial machines operating in their actual industrial setup has limited the efficacy of detection systems. Many different audio databases exist having collections of sounds from dummy (or real) systems operating in controlled environments but a collection of audio sounds from actual industrial machines is missing. Therefore, audio sound recordings of an Air compressor machine working in its natural industrial environment are presented. Only real sounds of an actual machine are captured. Synthetic mixing of sounds is avoided. Damaging the machine to create an anomalous state is avoided. Yet fourteen different unhealthy states are identified and their audio recordings are presented. Dataset with varied values of SNRs is also presented. Spectrograms are plotted and spectral shape parameter values of the developed corpus are calculated. The findings demonstrate the divergence in the developed database and its usefulness in building an effective AAD system for a real industrial machine.