Application and Evaluation of Machine Learning Tecniques for Process Control


Macori i Longas, Marc


Given the rising trends in industrial environments to digitalise data collection, allowing for a fast retrieval of data at high volumes, classic quality control methods to detect out-of-control (OOC) processes such as statistical process control (SPC) have seen increased limitations. Due to that this project, tests the viability of using alternative methods to monitor processes within the popular field of machine learning (ML).
To achieve this, code is developed to create a synthetic dataset which simulates observations in an industrial process, which can be in-control or out-of-control. This data is manipulated so observations from different out-of-control situations are contextualised and differentiated from observations where the process is in-control. Then two ML algorithms, that can tackle the problem of identifying out-of-control processes are chosen, and ML models using these algorithms are created, validated, and tuned. Finally, SPC methods are automated. Both, SPC methods and ML models, specifically isolation forest and support vector machine-based models, are used to predict whether a process is in-control or not within the same benchmark.
The results obtained show that the ML models can be adapted to identify OOC processes and are ideal substitutes to SPC methods, as both methods have similar measures of correct predictions with F1-score of 0.71 and 0.72 using SPC method variations and 0.79 and 0.80 using isolation forest based mode and SVM based model respectively. A few differences between the methods can be observed, however. More precisely ML models have a better balance between identifying OOC instances correctly and not labelling in-control instances as OOC. On the other hand, SPC methods are slightly better at detecting OOC processes but at the cost of wrongly labelling more in-control situations as OOC. When observing the predictions per type of OOC pattern, it is possible to see that SPC methods have an improved ability to detect trends and the support vector machine-based model generally has lower number of correct identifications of OOC patterns than the other methods.



Cuadros Margarit, Jordi
Fornells Herrera, Albert


IQS SE - Undergraduate Program in Chemical Engineering