Parameter Optimization to minimize Production Line Errors

Leveraging machine learning to maximize production output

In major industries minor errors can lead to significant decreases in production outputs. One of our customers is running a large packaging site of a company active in the food industry and was struggling to improve its production output. Even though all the classic lean techniques were already applied, they did not manage to stabilize and continuously increase their production output. Our customer already had 2 data scientists in-house, but they were mostly working on creating and updating reports, not on structural reports. Therefore, they got in touch with Data Factory to fulfill their main objective of stabilizing and increasing the production output with the help of machine learning for parameter optimization.

To get started with this exciting project, we started by gathering and analyzing data from the ERP (Enterprise Resource Planning), the MES (Manufacturing Execution System), employee and maintenance systems. Simultaneously, we interviewed key people in the production and maintenance teams. By analysing this data, we could determine that the largest part of the output losses were caused by incorrect filling weights and by minor line stops.

Subsequently we decided to implement 2 high-impact use cases: 1) Production parameter optimization for avoiding wrong filling weights and 2) Identification of drivers of minor line stops with machine learning. For the production parameter optimization, we trained an XGBoost model to identify the effect of different parameters depending on a set of input material characteristics and then ran a simulation to estimate the impact of an improved set of parameters. For the minor line stops, we experimented with different machine learning models. Eventually a linear model was selected as the performance was not much lower than the best one, and it allowed to explain the effect of the different drivers of line stops. After have identified the drivers of line stops, we performed specific deep-dive analyses to identify the root causes and brainstormed on solutions to tackle them. One of the examples was a new shift handover process, as we identified that issues often occurred around the time of shift handover.

Finally, the results of the machine learning parameter optimization were tested on the line and led to a 60% decrease in filling weight issues. A new instruction was launched and integrated in the process confirmation process to ensure adherence to the new standard settings. After implementing the improvements for avoiding line stops, the line stops decreased by 25%, which together with the optimal parameters led to an output increase of 8%.