Anomalous Sound Pattern Detection for Machine Health Monitoring

Anomaly detection using audio signals from industrial machines in the manufacturing industry has gained broad interest over the last few years. For example, predictive maintenance solutions utilize raw analog signals to identify trends and patterns. In a few scenarios, an engineer working in a factory setting can tell when a machine is behaving abnormally just by hearing unexpected sounds (e.g., the loudness of sound) that are well within the human perceivable frequency range (20 Hz - lowest pitch to 20 kHz - highest pitch) which are typically concentrated in a narrow range of frequencies and amplitudes. The human perception of the amplitude of a sound is its loudness. However, the audio signal in its raw form is not always the best representation of the important features (e.g., frequencies, amplitude, peaks). Additionally, the machine learning applications which rely on using traditional digital signal processing techniques (e.g., digital signal processors, chips) have a lot of dependency on subject matter experts to tune the system for a better performance. Thus, we investigate how the digital transformation of waveform signals from microphone sensors (e.g., Audio recordings of industrial pumps, valves, slide rails) into Spectrograms (A spectrogram is a voiceprint of a signal which expresses an audio signal as an image using different colors to indicate the amplitude or strength of each frequency.) can help to monitor machine health (e.g., anomaly classification). In the pre-processing phase, raw audio signals (.WAV format) from each machine are converted to Mel Spectrogram images using short-term Fourier transformation. Then, comparative study of image classification techniques using deep convolutional neural networks (CNN) with and without data augmentation, is conducted to classify images as normal or abnormal. The approach is evaluated using Malfunctioning industrial machine investigation and inspection dataset [1–3] (MIMII dataset). Results show that the neural network based models trained on the dataset with Mel Spectrogram transformation perform better than models trained on the raw dataset (i.e., sound samples without spectrogram conversion).