Acoustic Signal-based Leak Size Estimation for Electric Valves Using Deep Belief Network

Abstract

To achieve the balance of plant, industrial valves are extensively used for critical safety and control functions. Conventionally, the threshold and the visual observation method are used for valve health monitoring. However, these methods are slow. This study presents a systematic application of deep belief network (DBN) for fault size estimation in the DN50 electric gate valve. First, real acoustic signals representing the malfunctions are acquired. Secondly, the influence of the transmission path and background noise from other equipment are decoupled, using wavelet packet decomposition and reconstruction. Finally, three different DBN models are developed for valve internal leakage assessment, using the original signals, time-domain parameters and the decomposed wavelet packets. Evaluation results show that the model trained with the time-domain signals achieve the optimal result. The model also shows the capability to automatically extract the deep features from the signal, escaping the dependence on the conventional signal processing method and reducing the signal processing time. The application of DBN for size estimation also solves the slow convergence problems in the conventional multi-layer, backpropagation neural networks.