Detection of food contaminants with Microwave Sensing and Machine Learning

Today, food contamination due to foreign body is still a trouble for food manufacturers. First of all, because they have to guarantee a safe product to consumers. Secondly because a contaminant can damage the company reputation and lead to expensive recall campaigns. Finally, because more accurate foreign body detection systems allow the producers to gain important food certifications that could increase their incomes. For these reasons many techniques are currently adopted to solve this problem, such as mechanical filters, metal detectors, X-rays imaging and near infrared imaging. However, there are still some limitations: nonmetallic and low-density objects, like fragments of plastics, glass and wood, are not currently detectable; high penetration depth and low spacial resolution trade-off is still relevant; some methods are subjected to water attenuation. The good new is that Microwave Imaging Technology has the potential to overcome those problems in addition to other attractive characteristics. Indeed, it is non-destructive, contact-less, non-ionizing, real-time, cost-efficient and easy to operate. The goal of this thesis is to apply Machine Learning (ML) algorithms to a Microwave Imaging (MWI) system prototype to identify foreign objects in hazelnut-cocoa spread jars on the conveyor belt. Two kind of binary classifiers are employed: a Support Vector Machine (SVM) and a Multilayer Perceptron (MLP). The training is performed on two different static datasets. The first one consists of 1800 synthetic tomographic images. A 10-folds cross-validation (CV) accuracy of 99.167% is reached for the SVM and a 5-folds CV accuracy of 99.167% is achieved for the MLP, with an error of 0.278% and 1.111% over a test set of 360 samples, respectively. The purpose of this dataset is to validate the idea of applying ML to the foreign body detection problem in the hazelnut-cocoa cream jars with MWI. Since the results obtained with the first synthetic dataset were encouraging, a second dataset is created. It is composed by 2400 S12 scattering parameters of real measurements, acquired with an MWI system prototype. In this case the results are: 10-folds CV accuracy of 95.052% for the SVM and 5-folds CV accuracy of 95.833% for the MLP, with 6.04% of error over a test set of 480 samples, for both. To conclude, since both the best SVM and the best MLP give the same error rate on the test set, the latter is chosen for a faster hardware implementation. It is translated into an High Level Synthesis (HLS) C/C++ code and then synthetized for the Xilinx Zynq®-7000 SoC with Vivado HLS. This allows a large and quick design space exploration to satisfy the 100 ms latency requirement of this project, while optimizing area, power and throughput. The classification performances on the same real test set are confirmed. Regarding the future works, the main perspectives are to: enlarge the static real dataset, including also new types of intrusions, such as wood; train different classifiers, especially an ensemble; validate the classifiers with dynamic measurements; discover which is the optimal switching sequence of active antennas pairs during the transit of the jar below the arch to maximize the illumination of the target; validate the behavior of the system on different homogeneous food products, such as honey, yogurt, baby food, and on non-homogeneous ones, like chocolate spreads with hazelnut grains, to have a wide range of possible application scenarios.