Empirical Study of Machine Learning Methods for Predictive Maintenance of Industrial Compressors at Gas Processing Facilities in Edo State

Gas processing facilities in Edo State, South South Nigeria, depend on reciprocating and centrifugal compressors for gas compression operations, and unplanned compressor failures result in production shutdowns, safety risks, and significant financial losses. This study presents an empirical comparison of machine learning methods for predictive maintenance of industrial compressors using condition monitoring data from three gas processing facilities in Ologbo and Orhionmwon Local Government Areas. Vibration, pressure, temperature, and flow sensor data streams are collected from six compressor units over an 18-month period, including data from intervals preceding six documented compressor fault events encompassing valve failures, piston rod wear, and bearing degradation. Three machine learning models, a random forest classifier, a long short-term memory recurrent neural network, and an isolation forest anomaly detection model, are trained and evaluated using a time-series cross-validation methodology. Feature engineering from raw sensor data includes statistical features, spectral power ratios, and cross-sensor correlation coefficients. The LSTM network achieves the highest fault prediction accuracy, correctly identifying incipient faults up to 72 hours before occurrence in five of the six historical fault events, with a false alarm rate of 8.3 percent over the full 18-month dataset. The isolation forest anomaly detector provides earlier anomaly flagging at 96 hours before fault occurrence but at a higher false alarm rate of 21.4 percent. The study develops a practical implementation framework for deploying the recommended hybrid LSTM-anomaly detection pipeline at Nigerian gas processing facilities. Keywords: predictive maintenance, machine learning, compressor, LSTM, gas processing Nigeria