Data augmentation on convolutional neural networks to classify mechanical noise

Mechanical noise identification and classification are essential for automotive and machinery fault diagnosis. The scarcity of labelled audio data for noise-related mechanical issues has constrained the utilisation of complex, high-capacity machine learning models. In this research, the application of augmentation methods for labelled squeak and rattle datasets has proven that the accuracy of deep convolutional neural networks can be improved. Data augmentation has eliminated common machine learning issues such as overfitting observed in the models trained from a small dataset. The influence of different augmentation methods for the dataset was evaluated and compared based on classification accuracy. Different data augmentation methods have been tested to classify audio classes of different mechanical noises. The use of class-specific data augmentation leads to the development of more accurate machine learning models. Different combinations of data augmentations were investigated. This research showed that the new combined augmentation process could significantly improve the classifiers’ accuracy. The proposed combined data augmentation technique achieved the highest precision with the lowest error rate for both squeak and rattle datasets. The proposed method can be applied to any type of mechanical noise identification and classification.