Alicia Manglano

Data Analyst @ Digital Factory Vorarlberg / FH Vorarlberg

Alicia Manglano holds a MSc. in Information Systems with a major in Data Science. Her career in the field of data analytics started in 2016 as a data scientist at Kasatria Technologies in Kuala Lumpur. Until 2018, she worked for the Zumtobel Group in the business area, where she developed a predictive pricing tool based on dynamic market prices. In August 2018, she joined the Data Analytics & Intelligence team at the Digital Factory Vorarlberg. Her field of research includes anomaly detection; process mining, supervised and unsupervised machine learning.

Prediciting Unusal Testing Results in Manufacturing Using Anomaly Detection

The ability to detect and consequently prevent anomalies during production is crucial and directly concerns the efficiency of industrial processes and goods. This study aims to thoroughly analyze and ultimately apply commonly used supervised and unsupervised techniques to a real-word industrial problem, in which products are tested prior to delivery.

During production, each instrument is subjected to a variety of tests. If necessary, instrument parameters are calibrated during each of the testing stages. The data set currently available includes measurement values from one type of instrument over a three year time period. Despite one instrument has passed all test stages positively, it happens in rare cases that a final functional test is negative. Since each test is requiring significant resources and time, it is of utmost importance to predict the final test result as early as possible. This study focuses on the evaluation whether it is possible to forecast the final test result at any stage of the previously performed tests.

Since the industrial data generation process is complex and multi-sourced, several challenges arise when modeling the data. Although test bench related operations follow a certain sequence, some instruments undergo the same operation in loops or sometimes, previous operations are repeated in order to stay cautious in the results. Consequently, the nature of the given data is partially non-independent and identically-distributed (non-iid), which forces a significant reduction on the number of observations.

Results indicate that it is possible to model the final measurement based on the previous test stages. It is shown, that interpolation performs well whereas extrapolation beyond the range might lack accuracy.

Acknowledgment: This study was a joint work with Leica Geosystems, part of Hexagon.