Model-Based design of a Machine Learning algorithm for on-site food authenticity testing

Machine Learning (ML) algorithms are commonly used for a variety of classification tasks. Among them, food quality assessment, including adulteration, is gaining particular attention. While food fraud detection is typically accomplished in specialised laboratories, there is an increasing demand for innovative sensors combined with portable devices (portable e-tongues) with Internet of Things (IoT) features for performing on-site or point-of-need (PON) measurements to be shared on clouds. Portable e-tongues combined with smart features enabled by IoT and ML techniques have expanded the panorama towards smart analytical devices of interest to the agrifood sector and brought innovation to in-process analysis interfaces. It is well known that the ML training phase requires higher computational resources than those necessary for the classification phase, thus preventing its implementation on embedded systems. A possible solution is to conduct the training offline, e.g. on a Cloud Service, performing the classification process online on an embedded device. The coding effort of complex ML techniques can be supported by Model-Based design enabling the high-level description of the procedures in conjunction with the automatic code generation tools. A streamlined implementation phase allows for fast prototyping on different hardware platforms resulting in reduced time-tomarket. This paper presents a case study on the use of a Linear Discriminant Analysis (LDA) model trained to discriminate tomato purées of different varieties using an IoT e-tongue device developed in our labs. The trained ML model was Model-Based designed in the Simulink environment with HDL compatible blocks to increase code portability to different hardware architectures, and then implemented on an STM32 Nucleo L476RG board through automatic code generation. The classification task and the code timing performance were finally evaluated.