A Data-Driven, Statistical Feature-Based, Neural Network Method for Rotary Seal Prognostics

Abstract

Failure of the rotary seal is one of the foremost causes of breakdown in rotary machinery, and such a failure can be catastrophic, resulting in costly downtime and large expenses. Assessing the performance degradation of the rotary seal is very important for maintenance decision-making. Although significant progress has been made over the last 5 years to understand the degradation of seals using experimental verification and numerical simulation, there is a research gap on the data-driven-based tools and methods to assess the health condition of rotary seals. In this paper, we propose a data-driven-based performance degradation assessment approach to classify the running/health condition of rotary seals, which was not considered in the previous studies. Statistical time domain features are extracted from friction torque-based degradation signals collected from a rotary setup. Wrapper-based feature selection approach is used to select relevant features, with multilayer perceptron neural network utilized as a classification technique. To validate the proposed methodology, an accelerated aging and testing procedure is developed to capture the performance of rotary seals. The study findings indicate that multilayer perceptron (MLP) classifier built using features related to the amplitude of torque signal has a better classification accuracy for unseen data when compared with logistic regression and random forest classifiers.