Identifying abnormal operating conditions in industrial machinery represents a critical component of predictive maintenance strategies. Real-world implementations, however, are frequently characterized by unlabelled datasets, heterogeneous sensor sampling rates, and the non-linear nature of failure propagation. This study presents a framework for unsupervised anomaly detection applied to a compressor system monitored through multiple sensors, combining Ensemble Learning across three distinct architectural configurations. Experimental results, obtained from real operational data, indicate that, while individual models exhibit bias towards specific anomaly types, the intersection of ensemble detections yields a stable core of structural anomalies. A hierarchical classification of operational conditions is proposed, which is able to identify Core Anomalies, Global-Boundary Anomalies, and Local-Topological Anomalies, providing a reliable baseline for pseudo-labelling of unlabelled operational streams in newly instrumented assets. © IFIP International Federation for Information Processing 2027.

Cited by: 0