Anomalous Machine Sound Detection Based on Time Domain Gammatone Spectrogram Feature and IDNN Model

Anomalous sound detection (ASD) systems distinguish normal and abnormal machinery conditions on the basis of sound. While most ASD systems rely on the log Mel spectrogram, it lacks sufficient frequency resolution and performs suboptimally for rapidly changing sounds. Alternatively, the Gammatone spectrogram, extracted using a time domain Gammatone filterbank, offers enhanced spectrogram resolution. This study investigates the effectiveness of the time domain Gammatone spectrogram for ASD. To optimally learn the time domain Gammatone spectrogram features, an Interpolation Deep Neural Network (IDNN) model was proposed as the detection model. This model detects nonstationary sound frames highly reliably. An evaluation was conducted using MIMII dataset with area under receiver operating characteristic curve (ROC AUC) as the metric. Experimental results showed that our proposed method achieved ROC AUC of 92.5%, outperforming the log Mel spectrogram feature by 5.9 percentage points.