Predictive maintenance using estimation from time interval for butterfly valves

Abstract

• AI Long-term failure prediction of butterfly valves; implemented using RNN and MLP • AI trained using proposed time intervals required for a change in physical quantities • Valve leakage simulation conducted to create failure data applied to AI training • Proposed method improve prediction accuracy with reducing computational demand • MLP model using time intervals yields best performance of RUL prediction In predictive maintenance, accurate forecasting of the remaining useful life (RUL) for ship components is valuable because ships can only be serviced when docked in ports. Traditional data-driven approaches using recurrent neural networks (RNNs) involve iterative processes that predict one step at a time, increasing computational demands and limiting their application to low-specification systems. In this study, a direct prediction method based on the current quantity and time-interval data, which is the time required for a change in physical quantities, was developed to prevent iterative predictions and reduce computational demands. Numerical analysis simulated butterfly valve leakage in ship exhaust systems, with temperature variations in the exhaust pipe serving as detection metrics. Virtual data with different degradation tendencies were created based on these simulations to supplement limited operational data for neural network training. Long short-term memory (LSTM), gated recurrent units (GRU), and multilayer perceptron (MLP) models were employed and compared. The proposed method showed an average accuracy improvement of 21% and a computation time reduction by more than 50% when predictions were made for half-year remaining until failure. From a comprehensive perspective, the MLP model was found to be the most appropriate neural network for the proposed method.

Topics & Concepts