| Abstract [eng] |
Developments in electronic systems and advances of industrial equipment integration into enterprise databases means constantly increasing amount of accumulated data. Currently this data is often used to represent results, meaning they are used in reactive actions. The goal of this Final Degree Project is to use the historical data not only for representation or analysis of past results but turn them into a resource that could be used to determine production time of new products in the electronics manufacturing industry. The field of electronics manufacturing is chosen for analysis in this Final Degree Project, primarily in the company Kitron. A short presentation of manufacturing processes, employee time registration and current method for evaluating production time is given. Also noted is the method of determining the most financially important manufacturing processes. Examples of how production data and structure is given. Finally, a classification-based production time determining system is proposed as an improvement to the current company process of determining production time. Surface component mounting process is chosen as an object for classification because it has the greatest impact on the company’s financial results. In addition, the company database has the biggest amount of data on this process when compared to others. This data allows to develop and train classification methods. Because of various data reports, formatting etc., data aggregation and preprocessing had to be performed, compiling information on each production batch together with related data on products, actual production times, product structures and other characteristics that might have impacted production time. The unevenness of durations in the production of the same product in different batches was observed. Median value of the duration and filtering of the production batches using 1 sigma rule were introduced in addition to the average of the durations. Classes were assigned to aggregated data based on duration segments. Three stretches of duration at 30 s, 10 s and 5 s were evaluated. The durations of the classes are selected according to the purpose of the system - duration prediction as well as simple interpretation by the end user. Classification methods with nine data sets were tested using the MATLAB software package. The results with the best-performing classification method are presented for each data set. Most important features were evaluated. The products used to evaluate the duration determination system were not included in the training datasets of the classification methods. A comparison between the time provided by the company's calculator, the class proposed by the developed system and the actual production results is presented. The best performing classification method was Linear SVM, while using the 30 s time class. |
