Stable Fault Diagnosis Under Data Imbalance via Self-Supervised Learning in Industrial IoT

This study proposes a self-supervised fault diagnosis model for low-resource industrial IoT scenarios, addressing the challenges of scarce fault samples, complex data structures, and rapidly changing operating conditions. The model builds global temporal feature encoding through a Transformer architecture and achieves label-free feature learning using multi-view consistency constraints and sequence reconstruction tasks, which enhances the structural understanding of equipment operating patterns. To handle the high noise characteristics of industrial sensor data, the model incorporates a multi-scale correlation modeling mechanism that captures key dynamic variations across different time scales and improves the distinction between abnormal and normal states. A set of systematic sensitivity experiments is conducted, including analyses of network depth, sample ratios, and data distribution shifts, to evaluate model stability under complex data conditions. The results show that the proposed method achieves strong fault discrimination performance across multiple core metrics and remains stable even under severe data imbalance, demonstrating its ability to perform reliable fault identification in industrial IoT environments with limited labeled samples. Overall, the findings confirm the effectiveness of self-supervised feature learning in industrial applications and provide technical support for developing highly reliable equipment state monitoring systems.