Experiment on spectrogram feature based CNN for sound anomaly detection

This work proposes an anomaly detection method for sound signals observed from motor shaft rotation. From the recorded sound data, time frequency spectrogram image feature is extracted and convolutional neural network (CNN) technique is used to detect change in the sound data pattern. The proposed method is implemented on a prototype edge device and experiments on real-time motor sounds show that classification accuracy of anomaly sound is 83.23%, 80.56%, and 78.44% for three types of noises - babble, music, and factory, respectively at 6 dB SNR. According to this study, the CNN technique outperforms other classical support vector machine (SVM) and k-nearest neighbour (KNN) techniques by 16.62% and 24.15%, respectively.