Compressor Diagnostics by Means of Bagging-Based Machine Learning

Compressors are widely recognized for their versatility and extensive use across a broad spectrum of industrial applications. Given their critical role in process continuity, ensuring their reliable operation is paramount, as failures can lead to unplanned downtimes, safety hazards, and substantial production losses. Consequently, the early and accurate identification of both healthy and faulty operation is essential for effective maintenance planning. This study addresses this challenge by introducing a machine learning (ML) diagnostic framework that leverages bagging-based ensemble methods, by considering bearing temperature signal as a key health indicator. The proposed approach begins with the generation of bagged datasets to improve model robustness and generalization. Five ML models are trained and optimized by using data corresponding to healthy operational periods. Each model independently estimates the bearing temperature, which is subsequently used for threshold-based anomaly detection. The input variables of the models include suction and discharge pressures, oil head, rotational speed, and axial displacement. The overall health state of the compressor is determined by means of a majority voting mechanism across all models. The framework is validated by using real-world operational data collected from three compressors belonging to the same product family. Each dataset spans approximately eighteen months and is split into training data (labeled as healthy) and unlabeled inference data. The results demonstrate the capability of the proposed method in identifying faulty behavior under diverse operational and environmental conditions. In fact, the method achieves high accuracy in fault detection.