Spatiotemporal outlier detection methodologies for image-based process monitoring

During the past two decades, Statistical Process Control (SPC) has been continuously improved from monitoring individual data points to linear profiles to image data. Image and video sensors are now increasingly being deployed in complex systems due to rich information context that they can provide. As a result, image data plays an important role in process monitoring in different application domains such as manufacturing processes, food industries, health care, and structural health monitoring. Most of the existing process monitoring techniques fails to fully utilize the information of image data due to their complex data characteristics in both spatial and temporal domains. This dissertation proposes a spatiotemporal outlier detection methodology based on Partial Least Square (PLS) and two novel control statistics based on Area Delaunay Triangulation (ADT) and Area Total Squared Error (ATSE) of prediction error to improve the performance of image monitoring schemes. (PLS) regression, including Kernel and Sparse (PLS), and background reduction as an image processing technique are used to extract the important features of high-dimensional image data and estimate the in-control (IC) image of the product/part to provide the pixel fs color density errors between the in-control and sampled images. Then, the (squared) errors are connected using Delaunay triangulation and another geometrical algorithm to form the control statistic (ADT) and (ATSE), respectively, which leverage the geometry of data point to improve the outlier detection performance. A real case study at a paper product company with mass customization is used to demonstrate the performance of the proposed methodology under different out of control conditions in comparison with some of the existing methods in the literature.