Failure Prediction of Industrial Equipment Using Machine Learning

Abstract

There always exists an industrial equipment that must be operational and maintained at all costs. The necessity of such equipment depends on the application of the industry, the application varies from health care facilities, commercial buildings to industrial production lines. Predictive maintenance is a data-driven approach that aims to solve these problems by using Machine learning to determine degradation and failure of industrial equipment in advance. The paper presents an end-to-end logical approach to determine possible failures in advance of an industrial equipment that turns out to be chiller in this case. The data retrieved from The Building Management System (BMS) monitored the chiller components via sensors as a part of the Internet of Things (IoT) ecosystem that contains temperature, pressure, load, and alarm status. A unique relationship was discovered between system alarms and failure knowing that failure is also a function of the sensors. The approach can be divided into two stages, the first stage builds the alarms prediction model using input sensors that constitutes most of the work. The second stage uses the alarms prediction to classify chiller breakdown. The data is cleaned from garbage data and prepared in a proper format to be ingested into a pipeline that selects the best candidate features for each alarm using Recursive feature elimination. The selected features are combined with features defined by the physical system to be used in training multiple machine learning models with a prediction power of two-hours in advance. Recall is the main metric for this use-case that ranges from 84% up to 100% in some cases. The chiller failure classification model has achieved a 91.32% recall and 82.91% precision on the test set. The results validate the feasibility of the approach proposed.

Date of Award	Dec 2022
Original language	American English
Supervisor	Shakti Singh (Supervisor)