An Early Fault Detection Method of Rotating Machines Based on Unsupervised Sequence Segmentation Convolutional Neural Network

Abstract

Early fault detection (EFD) is vital for mechanical systems to reduce downtime and increase stability. The main challenge of EFD for rotating machines is to extract discriminative features from noisy signals to identify early faults. However, the lack of labels for the whole lifecycle data hinders the application of some powerful supervised deep learning methods in EFD. Besides, many EFD methods have to set a criterion manually, such as a threshold, to judge whether an early fault has occurred. To address these challenges, this article proposes a novel EFD method based on unsupervised sequence segmentation convolutional neural network (USSCNN). At first, frequency-domain features are extracted from raw signals and converted to 2-D gray images. Then, historical lifecycle data are labeled by USSCNN so that a CNN classifier can be trained with these labeled data. The deep features of the historical data learned by the CNN classifier are utilized to train the health index (HI) assessment model. The proposed method is tested on three bearing datasets. The results have shown that the proposed method can detect incipient faults earlier than the comparing methods with lower false alarms. Also, the HIs learned by the HI assessment model shown that the proposed method can extract discriminative features for EFD. More importantly, the proposed method can detect an early fault by the well-trained classifier, which avoids manual criterion-making. Results of comparison demonstrated the effectiveness and the robustness of the proposed method.