Self-Supervised Learning for Rare Fault Detection in Industrial IoT Systems with Limited Labelled Data

Rare faults in Industrial Internet of Things (IIoT) systems are difficult to detect because equipment usually operates under normal conditions, leaving few labelled examples of abnormal behaviour. Although industrial sensors continuously generate operational data, annotating these observations requires technical expertise and reliable maintenance records. This paper proposes a self-supervised learning framework that uses unlabelled multivariate sensor data to support rare fault detection with limited labelled samples. Sensor measurements are first cleaned, normalized, and divided into temporal windows. A compact temporal encoder is then pretrained through contrastive learning to capture relationships among sensor variables and patterns over time. Physically plausible augmentations generate complementary views of each window, allowing the encoder to learn representations without requiring fault annotations during pretraining. The pretrained encoder is subsequently fine-tuned with a lightweight classifier using a small labelled subset, while class-weighted learning addresses the imbalance between normal and faulty observations. The evaluation is designed to compare the framework with supervised learning from scratch, autoencoder-based anomaly detection, and conventional classifiers under different label budgets. Performance is assessed using precision–recall area under the curve, fault-class precision, recall, F1-score, false alarm rate, and inference latency. An ablation study examines the contribution of self-supervised pretraining, with data partitions separated by time or operating run to reduce information leakage. The study investigates whether representations learned from unlabelled operational data can improve label efficiency and rare fault sensitivity while maintaining computational requirements suitable for industrial monitoring.