Leavening control system based on machine learning techniques

Our research work describes how deep learning techniques can be applied with success to create a non-invasive method to control the dynamic density of dough rising during the fermentation process. This paper explains in detail the steps performed to train and apply on the field a Convolutional Neural Network (CNN) to monitor the leavening of a traditional Christmas Italian type of sweet bread made in Milan, called Panettone, that usually needs an accurate and supervised leavening process of around three days. One of our main goals was to prove how these CNNs and their learned inner representations could easily become the foundation for developing a remote-supervision framework capable to monitor the leavening process. Since the duration of this crucial phase is not exactly predictable, as it depends on many external factors, and usually takes place during the night, it makes sense to adopt a not supervised approach that is able to autonomously detect and notify to the bakery personnel (sending sms, email, whatsapp, etc.) when the density of dough is considered optimal, appropriate and ready to start the baking phases.Results demonstrated that a CNN based paradigm is more effective and more accurate than the current used empirical methods. The model converged and the average loss value was near to zero, even if the set of images and examples adopted to train and test the classifier was limited. Though applied in the bakery products context, the designed approach can be easily adapted to other monitoring tasks or industry domains, and its independence from User expert knowledge and specific artisanal skills can be considered one of the major advantages.