Incipient fault detection of sensors used in wastewater treatment plants based on deep dropout neural network

Abstract

Risk of sudden collapse of any industrial plant increases if small magnitude incipient faults are not detected at an early stage. The paper proposes an optimized Monte Carlo deep dropout neural network (MC-DDNN) to identify incipient faults of sensors installed in wastewater treatment plants using the historical dataset of the plant. Such faults usually remain invisible or are misinterpreted as noise signals due to their small magnitude but can be overcome by the proposed method. MC-DDNN easily identifies the incipient faults of sensors installed in a simulated wastewater treatment benchmark model as well as sensors installed in a real industrial plant. The tabulated results show the estimated probability of incipient fault in terms of percentage probability as detected by the MC-DDNN. The dissolved oxygen (DO) sensor incipient faults in benchmark simulation model (BSM2) are detected with probability ranging from 4.9% to 23.4% and DO, pH and mixed liquor suspended solids (MLSS) sensors of effluent treatment plant (ETP) are detected with probability ranging from 0.07% to 11.43%. This estimated probability of faults indicates the small magnitude of the faults and hence proves that the method is capable of identifying faults at an early stage to issue warnings for early maintenance of the plant.