How to improve process manufacturing productivity with real-world AI solutions

With Industry 4.0 comes the promise of leapfrogging in productivity, quality, and the overall return on your manufacturing investment. A key driver of those improvements is the large amount of data gathered from the sensors, actuators, and other connected IoT devices deployed on the shop floor. The availability of?IoT-driven,?real-world?data?combined?with the latest?AI?techniques?opens?the door for?practical?AI?use cases?to?help?solve real-world?manufacturing problems.

The new techniques,?specifically?Deep?Reinforcement?Learning-based AI agents,?allow AI?solutions?to?move?from?theory (research labs?and academia)?to?practical application on?shop floors.

Where can AI help improve process manufacturing?

There are many challenges associated with?optimizing process manufacturing. At a conceptual level,?one can think about three main areas where?each improvement can generate significant ROI.

Whether a process comprises steps such as extruding, melting, cutting, heating, cooling, forming, mixing, or more, typically, an AI agent can positively impact manufacturing by:?

1. Minimizing start-up time?and therefore reducing both lost material and production time loss
2. Ensuring the?end product?stays within specifications?and proactively adapt control system parameters (e.g.,?PLC set-points for various elements of the process) to avoid out-of-spec production drift.
3. Controlling?overall quality?to reduce defects in environments with changing conditions such as raw material variability, equipment wear and tear, and more

How can AI improve process manufacturing??

Traditionally training an AI agent required advanced data science knowledge (i.e.,?machine learning algorithms?from statistical to deep neural networks) and large amounts of human-verified training data.?It?is, in most cases, neither technically?nor?economically feasible in real life.

However,?it is now possible to develop effective AI agents by combining internal process expertise with the newest simulation and Deep Reinforcement Learning (DRL) techniques.

Toolchains,?such as?Microsoft Project Bonsai,?leverage?the experience of?subject matter experts (SMEs) to develop those?AI?agents without the need for in-depth data science.?From operators to research engineers, these SMEs?help?define?the best approaches to select the relevant data, simulate the process appropriately, and define the?DRL training?parameters.

DRL is a type of training that leverages simulators to train an AI agent that Project Bonsai will have automatically designed based on the system’s specificities: inputs, outputs, and complexity.

Who is already using this approach?

Although this approach is new, working with?the?Microsoft Autonomous Systems engineering team, Neal Analytics?has?already helped several customers design, train and deploy DRL-trained AI agents using Project Bonsai.

One example of?this process in action?is PepsiCo’s Cheetos manufacturing process?shown?below. In this example,?the process consists of two stages:?extrusion?and?baking.

The following video provides more high-level information about this project.

This video goes into more detail about its implementation

Finally, I spent more time on the concepts behind Bonsai, DRL, and how they apply to process manufacturing during this recent webinar

(Most of this article was originally published on Neal Analytics blog)