Manifold learning applied to various industrial applications

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


With the industrial world wanting to move toward an operational state of lights-out manufacturing (total plant control/automation), advanced data analytic tools are necessary to reach this goal by using smart learning structures such as artificial neural network (ANN) models. These models can learn plant dynamics through the extraction of machine data that can be acquired with today’s advanced data-acquisition systems. For a particular plant process this could be in the realm of hundreds of sensors on a multitude of machines. It begs the question, can a smart structure find any meaningful information on connected machine0 through machinei that can be optimized to improve the overall process to produce a better output quality. This idea is explored through various applications using a set of re duction algorithms known as Manifold Learning (ML). These tools reduce complex datasets down to a representation for time-proven data analytic methods. This is explored through simulation and practical exercises. These algorithms work with the geometry of a data surface and builds a distance matrix (D). This (D) is then used in further calculations to bring a higher di mensional dataset (D) down to lower dimensional representation (d) through a nonlinear mapping denoted by [F(.)]. It is all about lowering the complexity for a visual representation for interpretation, or for computational methods to find an input/output equation for prediction purposes. The first application is classification based. Snapshots of a process output are taken and a region-of-interest (ROI) is selected. The issue with this application is that each pixel in the image acts a dimensional coordinate. This research will attempt to make this analysis problem simplified through ML. The second application attempts to predict nonlinear gains from a reduced coordinate space from a two-link vertical manipulator. The noticeable difference here is that we are trying to predict a parameter (a regression prob lem). The third practical application is an implementation of the achieved AI smart unit. In testing, basic control was achieved with the developed generic framework.