Optimizing Cheetos, Speaking Viral, and Augmented Wind Farms

Making Cheetos consistently cheesy with Reinforcement Learning

PepsiCo uses Deep Reinforcement Learning to create an intelligent monitoring system that improves consistency in Cheetos manufacturing

Context

While it might be difficult to grasp for non-die-hard fans, the manufacturing procedure for products like Cheetos, Doritos, and Tostitos includes many complex individual inputs. The product specifications that range from cutting speed to water ratio all interact with each other. When the optimal combination is found, it allows the Frito-Lay division at PepsiCo to make the perfect final snack.

What's new

Wanting a more efficient way to consistently manufacture Cheetos, PepsiCo has teamed up with Microsoft to implement an AI-based monitoring solution. In fact, the Frito-Lay division wants to produce Cheetos with the proper attributes consistently to reduce waste and increase product quality.

The famous Cheetos puffs are made on an extruder, a machine also used to make tubing, tire treads, and wire coverings. Until now, line operators were manually selecting non-conforming puffs coming off the extruder. They relied on their expertise to identify defaults in shape, size, and puffiness. Moreover, they adjusted the extruder's input manually subsequent to the identification of non-conformity.

‍

‍

PepsiCo, in collaboration with the Microsoft Project Bonsai team, took an approach that utilizes Deep Reinforcement Learning. By creating a simulation space, the system translates specs of created Cheetos into positive and negative feedback for the Deep Learning algorithm. Using this technique, they are able to simulate a full day's production in 30 seconds.

The selected approach allows the use of specific sensors on the production line to oversee and quantify product quality almost continuously. That way, it is able to make setting modification recommendations to the operator or adjust the settings autonomously.

Why it matters

This implementation is a great example of autonomous system design as it reduces the delay between non-conformity and corrective action, allowing operators to focus on parts of the line that require human expertise.

The solution is not made to replace operators, but to assist their daily routine. Furthermore, they were able to help the development of the tool significantly. Their input was essential when programming safety rules for the system. For instance, they implemented a system that limits the rate of change to the machine's input as abrupt modifications can potentially damage or jam the machine. These safety controls are critical to designing such a solution, which makes the added input of the operator's expertise so valuable.

“The project brought together a powerful mix of technology, applied modelling skills and subject matter expertise to create innovation on the factory floor,” says Dylan Dias, who worked on the project.

What's next

PepsiCo claims that Cheetos are only the first product in a line of similar snacks that will benefit from Project Bonsai's AI algorithms.

“This is the future for process controls,” says Sean Eichenlaub, a senior principal engineer at PepsiCo.

Predicting potential COVID-19 mutations with NLP techniques

Researchers at MIT trained a neural network to understand infectious viral DNA grammar with NLP methods in order to predict virulent mutations

Context

The flu vaccine composition is reviewed each year as the influenza virus mutates. The scientific phenomenon behind this yearly variation is called "viral escape". The ability for viruses to mutate and evade our immune system to cause infection is an immense obstacle to vaccine development. In fact, it is the main reason curbing antibody-based vaccines for influenza, HIV, and SARS-CoV-2. What good is a vaccine that no longer works a year from now?

What's new

Researchers from MIT recently published a paper that proposes a Machine Learning technique to predict which mutations may lead to viral escape. Using an approach from Natural Language Processing using a model architecture that is called bidirectional Long Short Term Memory (LSTM), the authors trained their system on 45'000 variants of influenza, 60'000 of HIV, and 4'000 of SARS-CoV-2.

Their approach "is not unlike learning properties of natural language from large text corpora because languages such as English and Japanese use sequences of words to encode complex meanings and have complex rules (for example, grammar). To escape, a mutant virus must preserve infectivity and evolutionary fitness—it must obey a “grammar” of biological rules—and the mutant must no longer be recognized by the immune system, which is analogous to a change in the “meaning” or the “semantics” of the virus."

Their model's task is to fill in a missing amino acid in a sequence, much like a system that fills blank spots in a sentence. To do so, the model creates an embedding between sequences that represent their varying relationships. Following this procedure, they rank the generated results by evaluating the infectious capability of the filled-in sequence.

‍

‍

The results are quite impressive: 0.85 AUC in predicting infectious and antibody-evading SARS-CoV-2 variants. For HIV and influenza, it achieved 0.69 and 0.83 AUC respectively. For more details on the models, methods, and results, click here to see the original paper.

Why it matters

Using data-driven tools to predict dangerous viral mutations using only the virus' sequence is very impressive. In fact, viral mutations are usually discovered by hand when scientists regularly analyze DNA taken from patients. The current procedure has many flaws and is extremely time-consuming. The advantage of predicting harmful mutations speeds up this process and allows researchers to develop therapies in a preventive manner.

What's next

A lot of parallel research is being done to explore the similarities between human languages and the language of amino acids. The application of machine learning techniques to the field of bioinformatics is certainly not novel. However, the recent advances made are extremely promising and validate the immense potential for AI-driven solutions in pharmaceutical research.

Wind farms under your control

Wind energy companies are investing in AI technologies to forecast wind energy production and manage wind turbines remotely.

Context

The worldwide shift to renewable energy led by countries such as Sweden, Costa Rica, and Germany is based on replacing the use of fossil fuels and other non-renewable sources of energy with electricity coming from wind, solar, and geothermal sources. Unsurprisingly, global wind capacity is on the rise. Following a 19% year-on-year increase in installations, total wind capacity peaked over 651 GW in 2019. This change comes with some important challenges; the power production of weather-dependent sources is dynamic and the technology used for capturing energy from nature must be manufactured, monitored, and maintained at large scale.

‍

‍

“The way a lot of power markets work is you have to schedule your assets a day ahead,” stated Michael Terrell, the head of energy market strategy at Google, in a recent account in Forbes. “And you tend to get compensated higher when you do that than if you sell into the market real-time.”

What's new

New investments in technology leveraging AI are reshaping aspects of the wind energy industry. The advances are two-fold: (1) wind energy production forecasting using weather information and (2) condition monitoring systems for predictive maintenance in remote wind turbines.

First, Google and DeepMind are combining historical wind energy production with weather data to predict energy yield from the wind farms they source in the Central United States. Terrell states that they “use Machine Learning to take the weather data that are available publicly, actually forecast what we think the wind production will be the next day, and bid that wind into the day-ahead markets.” The results are quite astonishing: a 20% increase in revenue for wind farms.

While the models and methods used by Google remain confidential, there is extensive research on forecasting short term wind energy generation:

• Forecasting Short Term Wind Energy Generation using Machine Learning
• Wind power forecasting based on daily wind speed data using Machine Learning algorithms
• Machine Learning Techniques for Short-Term Forecasting of Wind Power Generation

Second, the use of condition monitoring systems (CMSs) is being extended to wind turbines. The AI methodology allows detecting fault before potential failure using data from vibrational sensors and operating conditions. This prediction is especially useful for remote turbines that are not easily accessible for maintenance. An AI system created by Bruel & Kjaer Vibro (B&K Vibro) of Darmstadt, Germany present the potential failure modes for a certain asset, each with a probability of certainty. The company has installed more than 25,000 data acquisition systems worldwide, 12,000 of which are monitored remotely. Their wealth of fault mode data has allowed them to train Machine Learning models to identify almost every imaginable potential failure mode.

For the engineers among you, here are some interesting papers on using Machine Learning to predict equipment health:

• An Industrial Case Study Using Vibration Data and Machine Learning to Predict Asset Health
• Performance degradation prediction of mechanical equipment based on optimized multi-kernel relevant vector machine and fuzzy information granulation
• Faults detection and failures prediction using vibration analysis
• The prediction of the residual life of electromechanical equipment based on the artificial neural network

Why it matters

As the global capacity of wind turbines grows by 10% year-on-year, the maximization of turbine production and availability plays a major role in the renewable energy market. The application of multiple and diverse Machine Learning solutions strengthens the business case for the shift to renewable energy, and wind power in particular.

What's next

“Our hope is that this kind of Machine Learning approach can drive further adoption of carbon-free energy on electric grids worldwide,” stated Sam Witherspoon, a DeepMind program manager, in a blog post. We remain determined that the democratization of AI will benefit society and climate action today, and tomorrow.