Self-Supervised Temporal-Structural Learning for Industrial Fault Diagnosis under Data Imbalance

Abstract

Industrial fault diagnosis in complex manufacturing systems is challenged by highly imbalanced data distributions and limited labeled samples. This study proposes a self-supervised temporal-structural learning framework that captures both temporal dynamics and inter-component dependencies in industrial sensor networks. The approach leverages contrastive learning to pretrain representations from unlabeled multivariate time-series data, followed by a graph-based encoder that models structural relationships among system components. The model is evaluated on a large-scale industrial IoT dataset comprising over 500 sensors and 8 million time-series samples collected from a real production line. Under extreme imbalance conditions (fault ratio < 2%), the proposed method achieves an F1-score of 0.81 and recall of 0.87, significantly outperforming supervised baselines such as CNN-LSTM (F1 0.69) and Random Forest (F1 0.62). The framework also demonstrates strong robustness under noise perturbations and missing data scenarios. These results indicate that self-supervised temporal-structural learning provides an effective and scalable solution for real-world industrial fault diagnosis, particularly in scenarios with sparse annotations and evolving system dynamics.