The extrusion and other machines at Mondi’s plant are large and complex, measuring up to 50 meters long and 15 meters high. Each machine is controlled by up to five programmable logic controllers (PLCs), which log temperature, pressure, velocity, and other performance parameters from the machine’s sensors. Each machine records 300–400 parameter values every minute, generating 7 gigabytes of data daily.Mondi faced several challenges in using this data for predictive maintenance. First, the plant personnel had limited experience with statistical analysis and machine learning. They needed to evaluate a variety of machine learning approaches to identify which produced the most accurate results for their data. They also needed to develop an application that presented the results clearly and immediately to machine operators. Lastly, they needed to package this application for continuous use in a production environment.

If a truck at an active site has a pump failure, Baker Hughes must immediately replace the truck to ensure continuous operation. Sending spare trucks to each site costs the company tens of millions of dollars in revenue that those trucks could generate if they were in active use at another site. The inability to accurately predict when valves and pumps will require maintenance underpins other costs. Too-frequent maintenance wastes effort and results in parts being replaced when they are still usable, while too-infrequent maintenance risks damaging pumps beyond repair.

Our hardware engineers were translating algorithms developed by algorithm engineers into HDL using written specifications, and without knowing exactly how the algorithms worked. If the FPGA implementation did not perform like our simulations, we never knew if the implementation or the algorithm was the problem. And even a small change to the algorithm meant rewriting most of the HDL.