Intelligent Machine Reliability with General Value Functions

  • Author / Creator
    Wong, Andy
  • This thesis investigates the use of general value functions for detecting anomalous behavior in machines. Identifying abnormal behavior is critical for ensuring the safety and reliability of any machine or industrial process. When the cause of these anomalies is due to accumulated wear on components over time, maintenance needs to be conducted before failure occurs. The goal of a condition monitoring system is to identify faults that precede machine failure, using only data collected during regular machine operation. Here, we develop a method of using general value functions for the semi-supervised learning problem of machine fault detection.

    This method of time series anomaly detection, named General Value Function Outlier Detection (GVFOD), is compared to nine existing methods of novelty detection, including eight multivariate techniques, and one method for time series data. We evaluate these algorithms on a machine failure dataset, collected from a robotic arm previously created at the University of Alberta. The dataset consists overwhelmingly of data collected under normal operation, in addition to five different types of artificially induced failure. It was found that GVFOD outperforms all other algorithms by mean F1-score when sufficient training data is provided, along with fault data for hyperparameter selection. At smaller training sizes, GVFOD performs similarly to multivariate outlier detection algorithms. When default hyperparameters are used, or when selected through expert knowledge without the use of fault data, it was found that GVFOD continues to outperform other algorithms with sufficient training data, whereas some other algorithms suffer.

    Furthermore, a simulation of the robot arm setup was developed using a mechanistic model, and parameter search was used to find unknown material properties and experimental conditions. Using this simulator, gradual failure data was generated and used to compare the performance of GVFOD, UDE (Unexpected Demon Error), and LOF (Local Outlier Factor). These results help us understand why GVFOD is superior to other methods for machine fault detection; the tracking behavior of GVFOD creates a boundary of normality tailored to the tail-end of training data. This allows GVFOD to better identify future normal data, while maintaining its ability to discriminate data arising from faulty operation.

    This work demonstrates the challenges of creating effective data-driven machine fault detection systems, and how GVFOD and reinforcement-learning methods are especially suitable for this task. This is a significant improvement upon existing methods of anomaly detection for industrial machine reliability, and contributes to the overall goal of improved system safety and operational efficiency. However, the methods presented in this thesis are generic, and are easily extensible to other fields where anomaly detection in time series data is desired.

  • Subjects / Keywords
  • Graduation date
    Spring 2021
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.