Predictive maintenance of compressor with LSTM-ADF for multi-sensor multi-condition interval forecasting

Abstract Wind-tunnel compressors and other large rotating machines operate under continuously varying loads and speeds, with dynamics and sensor relationships that evolve across operating regimes. In this environment, models must be not only accurate but also reliable and deployable, providing interpretable uncertainty bounds that operators can trust. We propose LSTM-ADF, a Long Short-Term Memory architecture for adaptive dynamic fusion, designed to meet these practical requirements in industrial health forecasting. The model learns inter-channel dependencies through a GraphGate fusion mechanism and adapts its internal representation using a FiLM conditioning pathway driven by power and rotational speed. It outputs multiple quantile forecasts to construct non-crossing prediction intervals and incorporates a dedicated DAP branch that produces per-channel risk scores suitable for online monitoring. Training is guided by physics-informed regularization so that axial displacement remains within plausible limits, radial vibration preserves physical symmetry, and uncertainty scales consistently with load–improving interpretability and operational trust without constraining expressiveness. We further construct an engineering-grade compressor dataset that captures realistic nonstationarity, transient behavior, and load variation through synchronized vibration and displacement channels sampled at one hertz. On this dataset, LSTM-ADF achieves competitive point accuracy while providing more calibrated, compact, and physically consistent interval predictions, along with earlier and more stable anomaly indications compared with LSTM, Transformer, and TCN baselines. These results highlight the model’s emphasis on reliability, interpretability, and deployable robustness, addressing the true needs of predictive maintenance and industrial monitoring.