An integrated LSTM-AM and SPRT method for fault early detection of forced-oxidation system in wet flue gas desulfurization

Safe and efficient operation of the forced-oxidation system is of importance to the wet flue gas desulfurization (WFGD). However, equipment and system failures are commonly found due to the long-time running, frequent blower switching, and heavy workload etc., especially after the ultra-low emission (ULE) renovation to meet strict emission standard in China. This work develops a fault early detection method to improve the predictive maintenance of the forced-oxidation system including blowers, pipes, and the slurry tank. A model based on long short-term memory (LSTM) network and attention mechanism (AM) is constructed to predict real-time operation parameters and compare with the measured values. Then the sequence probability ratio test (SPRT) is utilized to analyze the prediction-measurement residual and provide automatic and dynamic warning. All the data for model training and prediction are from the build-in distributed control system (DCS) without additional sensors. The LSTM-AM model proves to accurately predict time-dependent and highly relevant parameters. SPRT can sensitively perceive the fault-caused residual deviation while alleviating the noises. Industrial application to the cases in a 50 MW combined heat and power generation plant is then carried out. Results show that the bearing failure of the oxidation blower and branch pipes (immersed in the slurry tank) blockage can be forecast in advance when the incipient degradation occurs.