A Multi-Task Self-Supervised Fault Diagnosis Method for Rotating Machinery Under Limited Labeled Data

Fault diagnosis of rotating machinery based on deep learning has been widely applied due to its efficient feature extraction capability and has played an important role in ensuring the safe operation of industrial equipment. However, the existing deep learning methods heavily rely on large amounts of labeled data, and the application is constrained by the great difficulty and high cost associated with acquiring labeled data in industrial scenarios. Therefore, considering that existing methods often only use a single task for feature learning, this paper proposes a dual task self-supervised learning method for fault diagnosis called Global Shape Alignment with Local Discriminative Contrast (GSLDC), which significantly reduces the dependence on labeled data by collaboratively mining the global and local features of unlabeled data. This method designs two pretext tasks, namely, Global Shape Alignment (GS) and Local Discriminative Contrast (LDC), where the former builds a similarity matrix based on the Fast Dynamic Time Warping (FastDTW) algorithm to learn global features, and the latter utilizes triplet loss to enhance the fault discrimination ability for local features. The two tasks share the same encoder and are jointly optimized through a weighted loss function, while a simple classifier is used for fine-tuning to achieve fault classification. Experimental results demonstrate that the proposed method offers significant advantages for fault diagnosis under limited labeled data, with strong stability and generalization compared with state-of-the-art methods, thereby providing an efficient and low-cost solution for intelligent diagnosis of industrial equipment.