Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Unsupervised probabilistic and kernel regression methods for anomaly detection and parameter margin prediction of industrial design

Sivaramakrishnan, Jayaram (2021) Unsupervised probabilistic and kernel regression methods for anomaly detection and parameter margin prediction of industrial design. PhD thesis, Murdoch University.

[img]
Preview
PDF - Whole Thesis
Download (4MB) | Preview

Abstract

One of the significant challenges facing industrial plant design is ensuring the integrity of massive design datasets generated during the project execution. This work is motivated by personal experience of data integrity issues during projects caused by insufficient automation affecting the quality of deliverables. Therefore, this project sought automated solutions for detecting anomalies in industrial design data in the form of outliers. Several novel methods are proposed, based on the Hidden Markov Model (HMM) and a modified General Regression Neural Network called the Margin-Based GRNN (MB-GRNN) along with optimisation techniques that minimise computation time. An HMM was used for validating process plant tag numbers using a self-learning approach. Results from experimental data show that HMM performance is equivalent to that of a custom-made design rule checking algorithm. The choice of components in industrial design involves setting specific design parameters that typically must reside inside permissible ranges called “design margins”. The MB-GRNN has the ability to estimate these permissible margins directly from design data and indicate potential design errors resulting from the invalid choice of design parameters by identifying data points outside of the estimated margins as outliers. The extremal permissible margin boundaries are determined by “stretching out” the upper and lower GRNN surfaces using an iterative application of stretch factors (a second kernel weighting factor). The method creates a variable insensitive band surrounding the data cloud, interlinked with the normal regression function, providing upper and lower margin boundaries. These boundaries can then be used to determine outliers and to predict a range of permissible values of design parameters during design. This method was compared to Parzen-Windows and another proximity-based method. The MB-GRNN also benefits from a modified algorithm for estimating the smoothing parameter using a combination of clustering, k-nearest neighbour, and localised covariance matrix. The computation time for this difficult task is minimised using a new derivative-based method that was tested successfully using a range of root-finding functions. The efficacy of the MB-GRNN and associated optimisation techniques were verified using three multivariate design datasets. The experimentation shows that the regression-based outlier classification approach used in this project complements the existing Parzen density-based method. These methods used in combination are intended for implementation as a decision support system for checking the quality of industrial design data to help minimise design and implementation costs. It is expected that the unsupervised techniques presented in this research work will benefit from the ever-increasing automation of industrial design processes.

Item Type: Thesis (PhD)
Murdoch Affiliation(s): Engineering and Energy
Supervisor(s): Parlevliet, David, Lee, Gareth and Wong, Kevin
URI: http://researchrepository.murdoch.edu.au/id/eprint/62536
Item Control Page Item Control Page

Downloads

Downloads per month over past year