Hybrid GRU-VAE and Spectral Residual Anomaly Detection for OWC Predictive Maintenance: A Technical and Economic Assessment at Chu Lai Port

Abstract

The central coast of Vietnam holds substantial wave energy potential, in which Oscillating Water Column (OWC) systems deployed in this setting require robust condition-monitoring and predictive maintenance strategies to sustain reliable operation. The study proposes a serial hybrid anomaly detector in which a GRU-VAE learns the normal wave-driven behavior of the SCADA channels and Spectral Residual (SR) analysis is applied to the GRU-VAE reconstruction residual to expose short-term frequency anomalies. On a simulated test set of 7,481 time steps, with all thresholds and weights fixed on a validation set, the hybrid reached, on average over 3 simulated datasets, a Precision of 0.845, Recall of 0.864 and F1-score of 0.854 (AUC-ROC 0.962), comparable to, though slightly below, a standalone GRU-VAE (mean F1 0.868) and an LSTM autoencoder (mean F1 0.867); applying SR to the residual instead of the raw signals raised the AUC-ROC of the SR branch from 0.881 to 0.955. Taking a Precision of 0.91 and a Recall of 0.89 as the target detection performance, a 15-year life-cycle cost (LCC) model with a 9% annual discount rate yields an NPV of +USD 79,955 and an IRR of 14.76% (+USD 69,719 and 14.1% at the measured performance). Monte Carlo simulation indicated a 75.9% probability of a positive NPV under the stated input distributions, while Tornado analysis identified the failure frequency, the emergency repair cost and the recall of the detector as the main drivers of NPV. These findings indicate the potential technical and economic feasibility of integrating machine-learning-based predictive maintenance into OWC systems in Vietnam, under literature-based cost assumptions.