- 189 views
- 758 downloads
Detecting the Onset of Machine Failure Using Anomaly Detection Methods
-
- Author / Creator
- Riazi, Mohammad
-
When a machine or a component of a machine fails, corrective maintenance is performed to identify the cause of failure and decide on a repair mechanism to restore the machine to its normal working condition. However, because the machine has failed without any prior warning, a considerable amount of time for procuring and repairing the failed component is required. Since machines and their components generally degrade through time, i.e., start in a normal condition and progress to failure, and the time of failure is not known in prior, a maintenance strategy will need to be considered to minimize the downtime of the machine and the service(s) it provides. As such, time-based maintenance strategies are predominantly used for maintaining the healthy condition of machines and equipment through a regular maintenance schedule. This helps with the extension of the operational reliability of the machine, preventing potential catastrophic failure(s), and reducing time required for maintaining the equipment due to pre-planning. Nevertheless, substantial disadvantages such as expenses for hiring expert technicians, replacement of parts regardless of considerable remaining useful life, imminent failure(s) in between scheduled maintenance, and increased risk of failure due to improper diagnosis or intrusive repair mechanisms has triggered the interest of the research community in further investigating other strategies for maintaining the healthy condition of machines. Hence, prognostics and health management through condition-based maintenance is suggested as an alternative strategy. PHM is an enabling discipline consisting of technologies and methods to evaluate the reliability of a product in its actual life cycle to determine the development of failure and mitigate system risk. Sensor systems are needed for PHM for online monitoring of an equipment. This strategy is referred to as CBM. PHM can be implemented in several ways: Physics-based, knowledge-based and data-driven-based. We propose anomaly detection as a data-driven technique for the early detection of fault in a machine. Anomaly detection is described as the task of identifying observations with anomalous behaviour (a fault) that can cause systems to deviate from their normal operating conditions in an unacceptable way. For the purpose of this study, a belt-driven robot arm test platform is designed. The robot arm is conditioned on the torque that is required to move the arm forward and backward, simulating a door opening and closing operation. A number of failures are simulated and data is collected. Several anomaly detection methods namely, $k$-NN, LOF, ABOD, HBOS, isolation forest, one-class SVM, PCA, T*-framework and deep neural network-based models including feed forward and convolutional neural network auto-encoders are tested. Data for normal condition and several simulated failures, e.g., loose belt tension, high temperature, etc. is collected. The normal operating conditions of the arm is learnt by the anomaly detection methods through training on samples of the normal only class and a threshold is obtained. The learnt model is then used on unseen test data which includes samples of normal data mixed with samples of failure modes and an anomaly score is computed per observation. The scores are then compared to a previously computed threshold which determines the normal or anomalous label. The performance of each model is evaluated using the precision, recall, F1-score, area under the receiver operating characteristic curve metrics and the average time required to train and test a model. Our results show that, the onset of failure can indeed be detected with a very high precision and recall and with an average F1-score of over 90\% for majority of the algorithms. Moreover, we further investigate feature-based anomaly detection of the torque time series data with hand-crafted descriptive statistic features and automatic features extracted from the convolutional auto-encoder. The reduction of dimensions through manual feature engineering show a positive impact both on the allocation size of data and on the performance of the models in terms of accuracy, time to train and test, and the size of the fitted models. The robot arm dataset is made available to the research community and the results of our comparative assessment of anomaly detection algorithms brings a significant potential contribution to the PHM and the CBM research field.
-
- Graduation date
- Spring 2019
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.