Predictive maintenance using machine learning

Predictive Maintenance (PdM) is an essential pillar for Industry 4.0. PdM enables users to know in advance when a machine is likely to break down so that the required maintenance can be scheduled. The abovementioned involves different aspects of PdM maintenance such as condition monitoring and anoma...

Full description

Saved in:
Bibliographic Details
Main Author: Amal Roy Lerroy Ashwin
Other Authors: Arindam Basu
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2020
Subjects:
Online Access:https://hdl.handle.net/10356/140317
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Predictive Maintenance (PdM) is an essential pillar for Industry 4.0. PdM enables users to know in advance when a machine is likely to break down so that the required maintenance can be scheduled. The abovementioned involves different aspects of PdM maintenance such as condition monitoring and anomaly detection. Structural components of machines such as Rolling Element Bearings (REB) can be used as parameters to indicate machine breakdown by analysing the historical log of their vibration signals data. The availability of machine data has gained popularity in the use of Machine Learning (ML) to ease the statistical analysis of vibration signals of REBs and build a ML model to execute PdM. Countless researches have been done to improve the use of ML for PdM in view of Industry 4.0. However, not much has been addressed about a generalized approach of PdM while overcoming the challenge of inadequate amount of faulty data. Therefore, in this project, ML models utilized for popular maintenance strategies such as condition monitoring and anomaly detection were reviewed in order to build upon and identify a generalized approach for PdM. Dataset of 12 REBs provided by the Center for Intelligent Maintenance Systems (IMS), University of Cincinnati was retrieved from NASA prognostics database. As the dataset is that of a time-series data, time-domain features were extracted and utilized to build three relatively popular ML models. They are Support-Vector Machines (SVM) and Multi-layer perceptron (MLP) for multi-class classification to simulate condition monitoring while Autoencoders (AE) to simulate anomaly detection. Eventually, it was found that anomaly detection with AE was best suited for a generalized approach across bearings of different characteristics. An ensemble of AEs was implemented with ensemble learning to further improve the performance of anomaly detection. Furthermore, two layers of thresholds as well as two-level of measures were constructed to propose a method in identifying three types of machine condition with anomaly detection. The thresholds could also be used to detect an upcoming breakdown of a bearing. This makes it possible for machine diagnostics and prognostics with the use of ML. The entire project was conducted in an offline setting of a machine operation.