Predictive Maintenance of Actuated Quarter-Turn Valves Using Artificial Intelligence

This paper proposes a predictive maintenance method for actuated quarter-turn valves used in oil and gas applications and carbon capture systems. The identification and classification of the degradation process allow the management of the most critical parts by adopting specific strategies. In this way, it is possible to avoid service interruptions and reduce recovery times. Therefore, the objective of the work is to develop a classification approach based on artificial intelligence algorithms capable of detecting the aging stage of the valves by processing the stem vibration signals, measured with accelerometers, in the initial portion of an opening-closing cycle. Several machine learning techniques are presented and compared in the processing of the information extracted from an experimental test bench: Support Vector Machines, feed-forward Neural Networks and Classification Trees process the statistical characteristics of the measurements; furthermore, a neural network with complex values is used to classify the same measurements in the frequency domain. A data augmentation procedure is implemented to create a meaningful training dataset for these supervised learning algorithms. Moreover, additional test datasets are generated simulating different disturbing effects (increasing noise level, lowpass filtering effects, superimposing DC components). The results show the possibility of identifying valve degradation in different situations and offer a very accurate comparison of the proposed classification techniques in terms of robustness against simulated disturbances.