Machine fault diagnosis based on data reassembly for a deep neural network trained by imbalanced acoustic recordings

Abstract

For high-tech equipment, online monitoring and diagnostics play a critical role in ensuring production under normal conditions. While the machines are generally in normal condition, fault conditions are often difficult, if not impossible, to simulate. In this paper, we propose a data augmentation approach to address the challenge of machine fault diagnosis for the above-mentioned scenarios where the normal and abnormal acoustic data sets are imbalanced. To overcome this limitation, we propose a data reassembly approach to generate new signal samples by randomly combining data segments from different machine states. The resulting sample is annotated as a vector according to the proportion of the corresponding machine state. The training with data reassembly is performed for a deep neural network (DNN)-based classifier using the log Mel spectrum as the acoustic features. To validate the proposed approach, we conducted experiments using a rotor kit. Acoustic data associated with the normal and seven abnormal conditions were recorded using a four-microphone uniform circular array (UCA). The results showed that the classifier with data reassembly had superior diagnostic performance compared to the classifier without data reassembly. The proposed method is also compared to several data generation baselines in terms of accuracy and F1 score.