Poster Abstract: Minimizing Labeling Efforts for Fault Detection and Diagnosis

Abstract

We present a semi-supervised fault detection method that combinescontrastive and active learning techniques to efficiently analyze sensordata in industrial systems. It uses a transformer-based encoderwith rotary positional embeddings for self-supervised pre-training,which allows it to build structured representations that can be usedfor density-based clustering using DBSCAN. This method significantlyreduces the need for manual labeling, as it only requires 20%of the data to achieve high accuracy in clustering. It outperformsunsupervised alternatives and is scalable and adaptable to evolvingfault patterns. Its reduced manual intervention makes it a valuabletool for real-world industrial health monitoring.