Data-Driven Prescriptive Maintenance: Failure Prediction Using Ensemble Support Vector Classification for Optimal Process and Maintenance Scheduling

Abstract

Maintenance can improve the availability of aging production systems and prevent process safety incidents. However, because of system complexity, resource allocation is nontrivial. This research developed and applied a framework to obtain optimal future-failure aware and safety-conscious production and maintenance schedules. Ensembles of nonlinear support vector machine classification models were leveraged to predict the time and probability of future equipment failure from equipment condition data. Multiobjective optimization of expected profit and a safety metric was then used to determine optimal process and maintenance schedules. The results of this research were that the ensemble models had an average accuracy and an F1-score of 0.987, that the ensemble models were more accurate and sensitive than the individual classifiers by 3 percentage points, and that the Pareto-optimal process and maintenance schedules were obtained, providing alternative solutions to the decision maker. This research described optimal resource allocation to help improve safety and system effectiveness.

Data‐driven prescriptive maintenance toward fault‐tolerant multiparametric control

“…The failure prediction and scheduling models follow a similar paradigm to that of the authors' previous work to which researchers interested in implementing are kindly referred 16 . This subsection summarizes the essential aspects for brevity and links them to the other framework components.…”

“…The objective of failure prediction is to estimate when a piece of equipment will likely fail. The authors used five key steps to predict failure in their previous work: (1) preprocessing of diverse system data, involving statistical feature generation and Z ‐score normalization; (2) feature selection, involving recursive feature elimination; (3) splitting, to divide data into training and test sets; (4) model creation, using multiple nonlinear support vector classifiers; and (5) prediction, consolidating individual SVM models into ensemble models over rolling prediction horizons 16 . It is noted that the mpSAMADA framework is flexible for use with alternative failure prediction methods.…”

“…The data is first preprocessed via Z ‐transformation 9 . Wrapper feature selection is then performed on the transformed data via recursive feature elimination 16 . A random forest classification algorithm is leveraged for this due to its intrinsic ability to rank features based upon their importance 33 .…”

“…(3) splitting, to divide data into training and test sets; (4) model creation, using multiple nonlinear support vector classifiers; and (5) prediction, consolidating individual SVM models into ensemble models over rolling prediction horizons. 16 It is noted that the mpSAMADA framework is flexible for use with alternative failure prediction methods.…”

“…9 Wrapper feature selection is then performed on the transformed data via recursive feature elimination. 16 A random forest classification algorithm is leveraged for this due to its intrinsic ability to rank features based upon their importance. 33 The selected features are then used to filter the data and provide a smaller An assumption is made here that after fault detection, either a sensor fault magnitude estimator is available or that sensor measurements can be trusted to provide a measure of the fault magnitude.…”