Predictive modelling of dynamic compressors using machine learning and deep learning techniques

The present study employs a predictive modelling technique for monitoring the Oil Temperature (OT) and Bearing Temperature (BT) in a multi-stage centrifugal compressor system used for wind tunne applications. The RTD sensors were used to collect real-time data for three operational phases: unloaded, loaded and re-unloaded of compressors used to supply compressed air to wind tunnel facility. The derived data were used to develop ML models based on sensor characteristics, thermal behaviour and turbulence in heat transfer medium. For oil temperature, random forest model outperformed the other models (R2 = 0.9202, MAE = 0.2506, RMSE = 0.3399) due to a constant flow of the liquid medium. The deep learning models (CNN, LSTM, CNN+LSTM) did not perform well due to underfitting caused by insufficient temporal resolution in the OT data. In contrast, bearing temperature measured in turbulent oil medium exhibited a delayed thermal response. Here, hybrid CNN+LSTM model effectively captured both the spatial and sequential features, achieving near-perfect prediction (R2 = 0.9992, MAE = 0.0137, RMSE = 0.0285). It is envisaged that models explored in the current work will enable proactive maintenance of compressors, resulting in increased productivity of the 1.2 m and 0.6 m wind tunnel facilities, as well as help in effective inventory management for spares of compressors.