A bi-contrast self-supervised learning framework for enhancing multi-label classification in Industrial Internet of Things

Abstract

In the Industrial Internet of Things (IIoT), multi-label classification is challenging due to limited labeled data, class imbalance, and the necessity to consider temporal and spatial dependencies. We propose BiConED, a bi-contrast encoder–decoder self-supervised model integrating two contrasting methods: RAC employs an encoder–decoder with augmented data to capture temporal dependencies and boost information entropy, enhancing generalization under label scarcity. QuadC captures spatial dependencies across channels through convolutions on hidden vectors. Evaluated on the real-world industrial benchmark SKAB, BiConED improves feature extraction for underrepresented classes, achieving a 26% increase in F1 score, a 67.72% reduction in False Alarm Rate (FAR), and a 57.25% decrease in Missed Alarm Rate (MAR) compared to models without the proposed contrasts. Even with limited labeled data, BiConED maintains a FAR below 1% and recovers up to 85% of the F1 score without resampling, demonstrating its robustness in imbalanced IIoT environments.