By combining high-resolution IoT devices' measurements and manually collected maintenance logs, we used Optimal Classification Trees to predict welding gun failure in the next 8-hour shift with high true positive rate, while keeping false alarms low.

In a matter of seconds, the output tree recovered failure modes that took years of experience to operators and engineers to observe and identify. The engineers on the floor carefully validated it at each node and considered it "making a lot of sense and showing some ideas I had but could not convince others of!"

Because the model shows step by step what machine settings result in failures, the insights can be used to improve the manufacturing processes.

For instance, we identified that a drastic pressure build-up indicates impeding failure for certain classes of equipments. Engineers confirmed that this signal was physically meaningful, and planned on improving the pressure control to reduce the machine failure rate.