DA-VICReg: a data augmentation-free self-supervised learning approach for diesel engine fault diagnosis

Abstract

Self-supervised learning aims to extract useful representations from unlabeled data by maximizing the agreement between positive pairs. However, traditional self-supervised learning relies on carefully designed data augmentation methods to generate positive pairs. When dealing with 1D vibration signals, data augmentation prone to potentially compromise the fault information in the original signals. Therefore, this paper proposes a data augmentation-free self-supervised learning framework for diesel engine fault diagnosis called Domain Adaptation Variance Invariance Covariance Regularization (DA-VICReg). The DA-VICReg uses cyclic angular vibrations collected within the same time period as positive pairs and extracts useful features from unlabeled data using a loss function composed of three terms: Variance, Invariance, and Covariance. We found that when positive pairs originate from different operating conditions, such as varying speeds and torques, the model can develop feature extraction capabilities that remain unaffected by changes in operating conditions. In addition, a spatial pyramid pooling layer and a trilinear attention module are used to extract vibration features at different scales and focus on critical spatial locations and channels. Finally, the proposed approach was validated through experiments on two types of diesel engines, and a comparison with prominent self-supervised learning methods confirms the superiority of the proposed approach. In engineering practice, this method can utilize a large amount of signals stored in different time periods for self-supervised training and learn useful features for downstream fault diagnosis tasks.