Artificial intelligence (AI) and machine learning are changing the way we produce and process food. PPO uses both AI and machine learning in the Smart Imaging System, to improve the safety and quality of what we eat. AI and machine learning allow the system to identify and remove even hard-to-spot foreign material in-stream. They also allow the system to measure and respond to product quality measures like fat/lean ratio, pH, and freshness and even identify myopathies like woody breast.
PPO’s software development team does a lot of behind-the-scenes work to make our system work at line speed in your plant. In this article, we’ll explain a little bit about how we build our AI tools to help you deliver safer, higher-quality food to your customers.
Preparing Your AI Models
Before we ship the Smart Imaging System to your plant, our software team builds specific algorithms for your system, based on your products and your specific needs. To do this, our developers need to train our AI models. And training is like teaching a two-year-old — by showing lots of examples and reinforcing learning.
As an example, let’s say we want to teach a toddler the difference between cats and dogs. To teach them, we will start by showing them 1000 photos of cat faces and 1000 photos of dog faces. Each time we display the photo, they will have to tell us if it’s a cat or a dog. If they get it right, we will show them another picture; if they get it wrong, we correct them. Once they can sort through all the photos correctly, we can say they have learned the difference between the two animals.
Now, if we show the same toddler a full-body photo of a cat, they wouldn’t know if it’s a cat or a dog. And that’s because the toddler wasn’t trained on what cats and dogs look like from the side. So, we need to train them on the side profile of cats and dogs. If we wanted to give the child a better understanding of the difference between the two animals, we would also need to give them examples of both animals sitting, running, jumping and sleeping. The more scenarios we give the child, the more data they will have available to them to use when they see a new photo.
The same principles apply when we train PPO’s system. The more examples of your product that we can add into our models, the more reliable the system will be at detecting something it has never seen before.
Why Do We Need In-Plant Data?
Our developers do their best to simulate in-factory conditions with our in-house models. For example, we can replicate the temperature and product type (e.g. chicken breast, wings). We can also copy the product’s average length, weight, and height. Not only that, but we can use the exact foreign materials that you see in your plant.
Even though we can re-create some conditions, we can’t precisely copy them all. Some of the plant-specific conditions that are important for our models are:
- Fluctuations in temperature and humidity
- Product condition (e.g. presence of water, marinade, ice, blood spots or veins)
- Distribution of product on the belt
- Extreme specular reflections due to water pools in products
- Cleanliness of the belt
These are just some examples of conditions we can’t re-create perfectly at PPO. That means we can’t fully rely on our in-house model to perform perfectly in your plant.
Sometimes conditions change in the plant. Workers switch in and out as the shift changes. Product size might vary when the supplier changes. We don’t know these things until we see them in your plant. So, the more data we have from your plant, the better PPO’s AI can learn and perform.
How Do We Collect Data At Your Plant?
When PPO’s system arrives at your plant, it has only seen our in-house data. It hasn’t seen plant-specific conditions like the ones we mentioned above, so the AI won’t work perfectly right away. To improve the accuracy of our models, we’ll need to work with your team to capture as much in-factory data as possible. We call this the Commissioning process.
During the Commissioning process, we will first observe your product running through our system in your plant, on multiple days and at different times and shifts. We’ll look at when the data was captured and the supplier at that time. We will also analyze differences in the product and belt conditions. Did the product have more ice? Less ice? How long has it been since sanitation? Are there bits of dried product on the belt?
After we collect this information, our software team will further train our AI models, optimizing based on the differences they observed. Then they’ll continue to monitor the in-factory data and fine-tune it even further. As our team collects more data and feeds that data to our AI, the system’s in-plant performance will improve even further. Eventually, the AI’s ability to detect and assess will stabilize. It will no longer have to rely on PPO’s software team to monitor and adjust. At this stage, the system is ready to inspect your products in real-time in your plant.
Interested in learning more about artificial intelligence in the food industry? Check out our webinar on AI and machine learning, titled “What’s in Your Data?”