AbstractToday’s industry is highly automated and offers an opportunity to log a variety of data for different aspects of the manufacturing process. Research suggests that mining those data can streamline those processes. During the literature review, it became apparent that much research for quality, logistics and chemical process improvements had been done but to the author’s knowledge, none of this research addressed improvement suggestions for assembly and joining lines based on already available sensor signals.
The project was split up into the three steps data collection, preprocessing and analysis. The proposed framework requires logging of all of the equipment’s sensor data (i.e. proximity switches, lights sensors), which is not a function offered for industrial equipment. Therefore a nomenclature based algorithm was developed to parse the underlying PLC (programmable logic controller) program to extract the data points of interest automatically. These data points then are directly converted into OPC (Open Platform Communications) setup parameters which enable immediate logging.
First the individual events within the log needed to be clustered into cases using novel heuristic algorithms. Evaluation of the resulting cases yielded that the data was flawed. Research suggests that such data can either be repaired, ignored or analysed with fault-tolerant algorithms like the heuristic miner. Since repair and fault-tolerant algorithms bear the risk of additional uncertainty, the deletion of incomplete or flawed cases was chosen. Identification of the faulty cases was addressed by enhancing the data with expert-based engineering features and tags for a fraction of the cases. Machine Learning then was applied to determine the completeness of the remaining cases.
Next expert knowledge was encoded into heuristic algorithms which can pinpoint several process problems within the data. Above described framework was tested with real-life data originating from an automotive body shop and the potential for cycle time improvements of up to 19% was discovered. It was also found that one of the shortcomings of the suggested framework is that the analysis is based on presumed dependencies between the different events. The Process Mining domain suggests algorithms that enable the discovery of the real dependencies by analysing corresponding event logs. Experiments with above algorithms yielded unsatisfactory results which triggered the development of an improved, for industrial assembly and joining equipment suitable Process Discovery algorithm. This algorithm, contrary to the established algorithms, can discover highly accurate models with a minimum number of cases. Based on this finding, rules were formulated to detect the questionable dependencies within a case. Finally, a new process called ‘Interactive Trace Induction’ was introduced and tested to enable capturing the crucial cases needed to discover such highly accurate models.
|Date of Award||Jul 2020|
|Supervisor||Yanguo Jing (Supervisor) & Rahat Iqbal (Supervisor)|