How AI can help food manufacturers reduce waste and optimize yield
Reducing waste during food production can drive significant ROI for manufacturers. However, the first steps needed to optimize manufacturing processes, knowing which process can be optimized and how can be challenging.
With new AI-powered technologies, food manufacturers can programmatically build and deploy AI agents to help with this optimization process. These AI agents are trained in advanced simulators using Deep Reinforcement Learning (DRL) technology. Once deployed, the agents can then either augment the human operator’s capabilities by offering optimization advice in real-time or automate manual actions under human operator supervision.
Waste and the food industry
Some recent studies estimate that food waste at the manufacturing level can represent 5% of the total food production. That’s 5% of raw material, energy, and equipment wear and tear wasted. Even if a large portion of this waste can be recycled for animal feed, these are hard bottom-line dollars lost each year. And these figures do not even account for the additional carbon footprint manufacturing these 5% twice represent.
Waste, be it raw material (feed), semi-finished, or finished product, is an ongoing issue for food manufacturers. There is always a struggle between ensuring consistency in output quality and maximizing throughput. Experienced operators know how to tune their extruders, ovens, and vats to optimize output while staying within product specifications. But experience takes time to gain and relying on a select few operators makes it challenging to scale production.
There are also some external variables, such as raw material variability or environmental aspects from equipment wear and tear and plant climate conditions, all of which are uncontrollable. These can throw off even the most experienced operator whose ability to fix issues will be limited by how quickly they can notice and react. Furthermore, top-level operators are in limited supply. Their expertise is seldom extracted and codified for all operators to leverage. Even if this (mostly) tacit knowledge could be codified, it would still have to be taught and used consistently across shifts, manufacturing lines, and plants.
Can AI help reduce waste and optimize production?
Solutioning this cannot happen with old-fashioned command and control systems. There is too much variability and too many heterogeneous parameters, both at the system and environment levels. Even with the same equipment, each line will be slightly different. A one-size approach won’t fit all.
That makes deep learning-based AI, i.e., using Deep Neural Networks (DNNs), a natural fit — at least in theory. These AI models are robust when appropriately trained, adapted to highly non-linear situations, and can improve rapidly over time as they learn.
In theory, and with the correct training data, one could train a DNN-powered AI agent to act as a “meta-controller,”, i.e., an agent supervising process controllers such as Programmable Logic Controllers (PLC) or PID controllers. This AI agent would observe the process inputs and outputs, the controller parameters (including setpoints), and the relevant environmental variables (e.g., sound, humidity, temperature, etc.) in real-time. Based on this combination of inputs, the AI agent would then change the controller parameters (e.g., setpoints) to adapt to the current situation.
However, traditional Machine Learning (ML) training that uses scores of labeled data to train a DNN can’t be applied here. Time and resource constraints in the manufacturing industry means it is not possible to generate hundreds of thousands to millions of training data points to train a DNN. At least not in a financially viable or effective way.
Fortunately, an ML technique is gaining momentum and can help: Deep Reinforcement Learning (DRL). Although the concept of DRL to train DNNs is not new (it was first tested in the early 1990s on simple control systems), only recent developments make this approach a viable one for real-life production systems.
Reducing food waste and optimizing production with Microsoft Project Bonsai
Training a DRL on a real-life system is not a viable option. It will, at best, waste more resources (raw material, energy, and time). Most likely, it will be impossible to implement. It may break equipment, endanger operators, or take months to train using “real-life time”. If training an AI agent with DRL is not possible using a live system, what would be the next best viable option?
Step 1: Leveraging SMEs for Machine Teaching
Autonomous Systems built on Project Bonsai start with Machine Teaching (MT). MT harvests the knowledge from the process experts, such as the human operators running the extruder controls, to help design the solution. MT helps identify the key parameters and heuristics particular to a given production process. This knowledge is codified into the AI agent architecture, the DNN input and output parameters, and the design of DRL reward functions.
Step 2: Building the AI agent and simulator
The second step is to build this AI agent, aka “brain,” using Microsoft Project Bonsai. Project Bonsai is a comprehensive toolchain that enables manufacturers’ process engineers and AI experts, like the consultants at Neal Analytics, to work together to build this AI agent.
In parallel with the AI agent design, experts develop the process simulator. There are multiple approaches to building a simulator. Selecting the most appropriate one will depend on the use case.
Step 3: Training the brain
With the brain design, reward function definition, and the simulator ready, the team can start training the brain using DRL. Before deploying the brain on the real-life system, the team tests it using the simulator too.
Once deployed, the brain can now offer real-time advice to operators to help them maximize throughput and reduce waste by tuning the system as needed.
As shared by PepsiCo’s CEO Ramon Laguarta, referring to the Autonomous Systems deployment on PepsiCo’s Cheetos production line, this technology can positively impact real-life food manufacturing processes.
Wait! There is more!
In addition to the sheer value of waste reduction, DRL has many additional use cases for the food manufacturing industry.
For instance, the robustness, the ability to manage non-linear processes, and the flexibility of Autonomous Systems AI agents trained with DRL allow manufacturers to also leverage this technology for other critical use cases such as:
- Processes where capacity is constrained: This occurs whenever there is insufficient production output to meet demand. It may be due to multiple reasons such as manufacturing waste, sub-optimal yield, lack of highly qualified operators, etc. Using an AI agent trained with DRL, it is possible to increase these processes’ yield by positively impacting one or more of these factors.
- Energy savings: DRL learns complex and often unidentified connections between people’s behaviors, building thermodynamics, and environmental variables such as people’s flow in a specific part of a building at certain hours, building insulations or lack thereof, and building sun exposure. This allows Autonomous Systems AI agents to reduce energy consumption by smartly adapting heating and cooling schedules and locations.
- Replacement of manual operation: By integrating top operators’ experience and tacit knowledge in these AI agents, manufacturers can efficiently convert tasks that could not be previously automated into automated ones.
Those examples are only a small subset of DRL applications for food manufacturers. DRL is widely applicable to most production processes. It’s an exciting new option to help manufacturers optimize yield and reduce waste across a variety of processes.
Want to learn more about DRL and Autonomous Systems? Contact one of our experts today to see how you can leverage the latest AI techniques in manufacturing.
- An overview on Autonomous Systems technology
- More about our Production Yield Optimization solutions
- Learn how PepsiCo makes the perfect Cheetos with the help of Autonomous Systems
- Unlocking the potential of AI in manufacturing with MT and DRL