Machine learning techniques for microwave food contamination detection

Ensuring the safe consumption of suitable food products is a top priority for food companies. There are several standard prevention systems that identify and eliminate contaminated products such as x-rays or infrared systems. Such systems are not reliable enough as they cannot detect low density materials, have low depth of penetration and x-rays can be harmful to operators. An emerging technique in the food world is Microwave Imaging Technology, which allows to overcome these problems and can detect low density materials. In this thesis, a Machine Learning algorithm is designed and implemented on a Microwave Sensing system that is able to detect foreign bodies in jars of cocoa-hazelnut spreadable cream during their passage on a conveyor belt at a speed of 30 and 50 cm/s. The system consists of a convey belt on which 6 antennas are positioned to acquire the signals when a jar passes under them, a network vector analyzer that acquires the signal from the antennas and processes it directly or transmits it to a microcontroller which processes it and detect the contaminat. The binary classifier based on the Multilayer Perceptron family has been trained with seven different datasets. A dataset acquired in a laboratory environment to find the hyperparameters that gave the best results, four other datasets subsequently acquired by the system installed in the company, two of which with canola oil and two with spreadable hazelnut cream, each one of them at two different speeds, 30 cm/s and 50 cm/s. Finally, the last two datasets are created by joining the datasets at two different speeds of the same material; the latter were created to verify the possibility of implementing a single algorithm that detects contamines regardless of the speed of the conveyor belt. The results obtained were very promising, using the best network found, with the first dataset an accuracy of 99.76% was achieved while using the dataset of the two-speed spreadable hazelnut cream we reached 99.91%. The network was loaded by converting it into High Level Syntesis C code and synthesized for two different systems in order to study the inference that gave the best results: 1) on the network analyzer computer whose output commands a microcontroller that signals the presence of contaminants. 2) on a microcontroller that receives data from the network analyzer. The best inference processing times were found in the case of the microcontroller, reaching an average of 10 ms per acquisition as opposed to the 13 ms of the network analyzer. Despite this, the performance bottleneck is in the communication of the data received by the Network Vector Analyzer, using a UART communication at 115200 bit/s and having 660 data in floating point per acquisition to be transmitted, the minimum transmission time will be 183 ms. Potential future improvements will be an expansion of the datasets with different materials and different contaminants, in order to have a greater number of contaminants that can be detected on different products. In order to reduce the communication times, a data acquisition system could be used that is able to manage SPI or I2C protocols which would reduce the communication times to the same order of magnitude as those of inference. In addition, the system could be made cheaper by replacing the Network Vector Analyzer and assigning the data acquisition task to the microcontroller so as to definitively eliminate communication times.