Data analysis and quality prediction

Situation

Our client, a production factory of a major Tier 2 automotive supplier, understands the potential impact of digitization and employing artificial intelligence in its production processes. They decided to engage with a specialized supplier to guide an internal team through an AI pilot project, which was designed to skill up the internal data team and show the potential of AI.

After selecting Cognexa as their AI partner and conducting a discovery phase, we have identified a solution concept which would be used by the assembly department.

Solution

Solution would grant the assembly department the possibility to predict if the produced unit is scrap before it is fully finished. This would lead to lower costs because the scrap unit does not have to run any further tests and it will also give clients the opportunity to investigate which settings and test results lead to scraps in production. This can be used to decrease the whole scrap ratio.

Solution used as input data results from the previous testing stations like, pressure measurements, brush cycles, etc.. Our solution would calculate if the unit will be scrapped in the next testing station or not. 

Our solution was fully Dockerized and cloud compatible. Core of our solution was done in Python and its libraries which were also used to automatically communicate with clients’ Amazon Redshift database. Final solution would be possible to deploy on the AWS cloud. Results could be exported back into clients database with possibility to incorporate them in clients inhouse visualization tools.

Final solution was using the Random forests for the classification if the unit is scrap. We were working with a dataset with two years of historical data. Dataset was composed from two large groups:

• Assembly data (results of the tests, measurements from the production line, additional features, production line settings, etc.)
• Energy consumption
• Quality data

For measurement of the model performance we chose the Mathews correlation coefficient. We decided for this metric because of the large imbalance of the dataset (low scrap ratio).

Result

All of the process was done on the clients premises in the cloud solution. Smooth work with Amazon Web Services and Amazon Redshift. Our solution would enable to lower the scrap ration. Lower scrap ration will be reflected in higher quality of production and reduced expenses related to scrap refunds (claims). 

Using much of the same data and code, we also implemented a solution for energy costs optimization for the client.