Sound Based DC Motor Classification by a Convolution Neural Network
Dejan Ćirić, Marko Janković, Miljan Miletić · 2022 57th International Scientific Conference on Information, Communication and Energy Systems and Technologies (ICEST) · 2022
Important properties of industrial products and their conditions are embedded in sound they generate. Thus, sound can be used to extract those properties including the product quality or presence of certain failures or faults. This can be done by applying an adequate classifier based on a deep neural network (DNN). Among a number of different DNNs, one of the most often used solutions in the audio domain providing the best and most reliable results are convolutional neural networks (CNNs). Here, such a CNN classifier for making distinction between non-faulty and faulty DC motors based on sound they generate is developed. The input to the classifier is a mel-spectrogram obtained after mapping DC motor audio signals into this image. Using only 668 labeled audio samples, the accuracy of about 87% is achieved.