The Business Problem

A manufacturing organization faced challenges with minimizing losses in their production line. Optimal control of a production line is a challenging problem. Traditional PID controllers monitor the process variable of interest and adjust a control variable to maintain the process variable as close as possible to the desired setpoint. The amount of control they exert is proportional to the error between the process variable and the setpoint (P gain), the rate of change of that error (D gain), and the accumulation of the error over time (I gain).

PID controllers often lead to oscillations in the process variable and can't model the physics connecting control variables to process variables. If variables change on the production line, traditional controllers will only react once the error has built up.

The INM Consulting Approach

With Reinforcement Learning (RL), we developed a system that can monitor all variables on the production line simultaneously (even hundreds of them), train a model to learn which variables drive the process variable of interest and in what exact way, and then use that model to optimally control the process variable by setting the values of multiple control variables at once.

Reinforcement Learning uses the same AI technology as autonomous driving, but instead of driving a car, it drives the production line, making optimal decisions in real-time to achieve the desired outcome.

Implementation Details

• Parsed real-time measurements from the production line (temperatures, pressures, setpoints, process variables)
• Created a digital twin of the production line
• Trained a Reinforcement Learning controller to take actions that minimize the loss function
• Deployed the controller on the production line

Technologies Used