Active Foundational Models for Fault Diagnosis of Electrical Motors

Fault detection and diagnosis of electrical motors are critical for the safe operation of industrial systems. Early detection allows for corrective measures that mitigate fault severity. Current data-driven deep learning methods for machine fault diagnosis rely on substantial labeled data, which is often costly to obtain. This study addresses the underutilization of vast amounts of unlabeled condition monitoring data by introducing a foundational model-based Active Learning framework. This framework leverages a minimal number of informative labeled samples while utilizing extensive unlabeled data through a combination of Active Learning and Contrastive Self-Supervised Learning techniques. The approach incorporates a transformer network backbone that is trained using a novel nearest-neighbor contrastive self-supervised learning method, enhancing the learning of representations from raw, unlabeled vibration data. The backbone can be fine-tuned for various downstream tasks across different machines. The proposed methodology's effectiveness is demonstrated through fine-tuning on three distinct machine-bearing fault datasets. Experimental results reveal that the framework outperforms existing state-of-the-art fault diagnosis methods while requiring fewer labeled samples.