Electronic Thesis and Dissertation Repository


Master of Engineering Science


Electrical and Computer Engineering


Capretz, Miriam


Nowadays, asset maintenance plays a vital role in companies and countries because if an asset breaks, it can cause long downtimes that may affect companies' production costs or the comfort level of building occupants. Predictive Maintenance (PdM) performs maintenance based on the asset's health status indicators. Sensors can measure an unusual pattern of these indicators, such as an increased motor's vibration level or higher energy consumption, and, in most cases, failures are preceded by an unusual pattern of these measurements. The increasing number of sensors are generating a massive amount of data. Machine Learning (ML) techniques can capture and learn patterns from the data and create models to evaluate the required health indicators. These health indicators are obtained by sensors, which are becoming more present every day. Convolutional Neural Network (CNN) is a Machine Learning technique capable of extracting data representation. This ability means that less feature engineering needs to be done when compared with conventional ML techniques. This thesis presents a CNN framework to tackle assets predictive maintenance problem and a method to transform 1-dimensional (1-D) data into an image-like representation (2-D). A data transformation step is very important to make the use of CNN feasible. To evaluate the proposed framework two datasets were obtained from fans, with distinct electrical pattern, from the CMLP building at Western University. The results presented by the CNN-PdM framework showed that the combination of CNN with the proposed data transformation method outperformed conventional machine learning techniques. The model created by the CNN-PdM framework achieved accuracy rates as high as 0.98 for one of the datasets and 0.95 for the other.

DOC060519.pdf (199 kB)