The manufacturing and processing industries are continuously improving the digital infrastructure to improve the degree of digital control, automation, and data collection for analysis in their facilities. These improvements result in the increased availability of digital records which give more insights on the prevailing conditions for each workpiece processed through the manufacturing line including timestamps, the sequence of manufacturing steps passed by each workpiece, and various physical parameters recorded by various sensors throughout the manufacturing line. The identification and root-cause analysis of quality defects on the end product is an inherent priority in such industries. With the recent improvements of digitization, smart data analysis using advanced machine learning methods allows for the quick identification of facts that help in the disintegration of complex production-defect relations. The main advantages of such an analysis are in the increase of production efficiency, reduction in the defect analysis time, and reduction in the requirement of process knowledge experts to identity these relations.
Generally, machine learning (ML) refers to an application that learns specific distributions and relations from a specific dataset. These machine learning algorithms fall into two major categories namely supervised and unsupervised machine learning algorithms. Supervised learning models learn various feature distributions of a structured or labeled dataset by iterative training, whereas unsupervised learning models learn the feature space by identifying characteristics of the original dataset.
Through this thesis various machine learning approaches are explored, to identify a precise framework using a discriminative unsupervised machine learning (one-off application) model based on Hebbian learning principle to extract information from quality inspection datasets to help identify possible factors that contribute towards quality defects on the end workpieces. Such results can help manufacturing facilities in the identification of time periods where quantized sets of defective workpieces that affects the defect ratio predominantly are identified and more in-depth analysis of these clusters to identify embedded periodicities in a certain type of defect clusters. This information can aid the manufacturing plant operator significantly in making observations and changes to various process steps to optimize the manufacturing techniques and reduce the overall defect rate.
Keywords: Big Data, IIoT, KDD, Time-Series Clustering, Neural Gas, CFD-periodicity analysis, Local Regression