AI-Learning to Roast

ORCHESTRA Extreme Roasting Control is the TMR software that has been identified as a first-rate starting point to experiment Machine Learning Intelligence. The project aims to enhance the reliability and productivity of the roasting process in industrial sized equipment with the use of machine learning modelling, which is able to predict the product profiling and processing conditions, optimising productivity and energy saving.

A quality roasted coffee is the result of an appropriate heat treatment of the raw beans combined with the customisation of the final taste. It requires in-depth knowledge of the product and the technologies to properly manage its transformation and really impact coffee’s chemical reactions.

It might be taken for granted, but now more than ever, precision in determining the consistency of the aroma of a single coffee package has become the paradigm for developing smart solutions for superior productivity in the process upstream. Certainly, coffee roasting is one of the most decisive steps in the whole process.

Especially during roasting, some raw product characteristics such as humidity, bean density and the maturation age of the harvest, can dramatically affect roasting and the roasted product quality and ultimately reliability. The green beans contain a significant amount of amino acids, proteins, sugars, oils, carbohydrates and caffeine.

During roasting, changes in the physical and chemical properties of these non-volatile compounds are responsible for the quality indicators in a cup a coffee, such as the colour, taste, aroma, body, viscosity, density and odours.

Thus, in-depth knowledge of the product and its transformations means you can make an accurate evaluation of the steps required to reach the desired quality over time, batch by batch.

Roasting control is fundamental in trying to determine and establish from the beginning what the final characteristics of the finished product will be, in terms of taste, olfactory, organoleptic and even physical characteristics, that is, the appearance of the product. Nowadays, each coffee company requires accurate and detailed flavour profiling of their product, in order to enjoy good sales of a product while developing new ones.

While quality assessment is crucial to meet consumer requests and to offer products with high standards, traditional methods to evaluate the beverage quality requires time-consuming and several costly practices.

This does not allow results in real time and usually requires an expensive waste of good quality coffee and skilled personnel to operate the machine.

IMA maintains that the right path to reach higher standards in machine intelligence is the one that enables the experience and knowledge of the manufacturer to be embedded into the machine itself.

Machine learning is a branch of artificial intelligence, a computer-based system that may be trained to find patterns among top-ranked data, comparing dozens of models the software has already memorized. It is based on an auto-tuning model that acts as surrogate for a hand-crafted algorithm, able to classify or predict specific parameters and improve its performance by feeding new data. The result is used to find insights and predictions and solvemore problems faster, more accurately, objectively, cost- and time-effective way and with higher safety levels.

In the last decade, we are seeing automation enter every step of the coffee supply chain, developed to save time as well as minimise human error, anticipating any unexpected situations and helping to obtain the consistency required. The use of Machine Learning has been increasing in recent years also in the food and beverage industry, because product quality statistically decreases a few years following its launch.

One of the key benefits that the “Ai-learning to roast” project aims to obtain is that of addressing the increasing need for targeted coffee profiles, maintaining quality in the long term, while dramatically improving productivity, energy saving and reducing environmental impact.

Today, ORCHESTRA enables full personalisation of the recipes by previously creating the roasting profiles and by setting the time and bean temperature as independent variables. It is based on a complex and articulated cascade control scheme, that allows the creation of precise constraints along the roasting phases. The project on Machine Learning applied to ORCHESTRA is expected to increase the reliability, especially on the setting-up of new non-conventional roasting profiles and during frequent recipe changes, anticipating unexpected situations and minimising the fine-tuning operations made by highly specialised technicians.

The system will be able to automatically recognise the patterns to be adopted considering the coffee processing parameters and the machine’s configuration. Regardless of the machine’s capacity per batch, the airflow rate, the afterburner power, the presence of a catalyser, the fan pressure, the inlet hot air and emission air temperatures needed, the ML technology recognises in a short time the machine’s behaviour and applies the pre-set profile perfectly, adjusting the parameters consequently.

In line with IMA’s vision of a Smart Machine, the added value of the project will result in reduced time-to-commissioning and, therefore, reduced time-to-market. Anticipating a possible problem on the setting up of a new product recipe can drastically reduce unplanned production stoppages, while establishing a predictive maintenance plan.

The most complex section of the project aims to let ORCHESTRA run as a predictive tool, the smartest intelligent roasting assistant.

Research is expected to integrate the software with the following skills:

* ability to handle the roasting phases in the time needed, adjusting the parameters consequently;

* ability to provide useful information on machine health and life expectancy;

* ability to sort out prognosis services for troubleshooting and predictive maintenance plans.

In addition, accurate tracking of various inputs will lead to easier data management optimisation.

Today, one of the strengths of the Petroncini TMR model is that it has substantial benefits regarding energy savings and a limited impact on the environment in terms of CO2 and VOC emissions, up to 40% less compared to open cycle roasters present on the market. Through the optimisation of the entire process, achieved thanks to better adjustment of all the roasting parameters, the aimed result is to significantly reduce the carbon footprint.