A Proximity-Based Unsupervised Feature Learning Framework for Rotating Machinery Sensor Data

With the advancement of health monitoring technologies in the era of smart industry, vast amounts of sensor data are continuously collected from rotating machinery. However, labeling this data remains a major bottleneck in industrial applications. This paper proposes a novel unsupervised learning framework for fault diagnosis, based on the assumption that sensor signals within adjacent time intervals exhibit high similarity in health states. By maximizing the proximity between non-overlapping, temporally adjacent sample segments, the proposed method effectively extracts discriminative features without requiring knowledge of the number of fault types. The approach is evaluated on three public benchmark datasets through unsupervised clustering and label matching. Experimental results show that the method significantly outperforms existing unsupervised techniques and achieves accurate label alignment without expert intervention.