A Machine-Learning Approach to Enhanced Food Safety via In-Line Microwave Sensing

Physical contaminants continue to be a major challenge for the food industry. These intrusions elude controls and end up in consumer’s hands and bodies, posing a serious health issue and a high level of media coverage: accidental ingestion of foreign bodies can cause choking and seriously damage the digestive tract. Precautionary methods and systems already in use, such as X-rays, metal detectors and near-infrared imaging, may fail to detect particular kinds of materials, especially low-density ones, and the increase of plastic employment in industries, together with the rise of automation in production facilities, may cause additional foreign bodies contaminations. This thesis presents a novel detection principle based on microwave-sensing technology and explores deeply the field of machine learning (ML) in search of a fruitful combination between these two. The former is focused on exploiting the new detection principle, the dielectric contrast between the potential intrusions and the surrounding matter. Even if being really low, it generates electromagnetic signal alterations that can be sensed by dedicated instruments and points out the presence of unwanted intrusions. It works on packaged products so that no further steps in production may accidentally contaminate products, as they get inspected just before going out of the production facilities The latter receives the dataset of measurements composed by microwave-acquired signals and introduces it to a set of ML algorithms and classifiers to identify contaminated samples. Thus, this research scope gets divided into two main goals: 1. To implement a binary classifier to discern clean from contaminated samples. 2. To implement a multi-class classifier capable of differentiating among millimetric-sized intrusions made of plastic, glass or wood, the classes of materials unlikely to be spotted by existing inspection devices. The first objective investigates different model architectures while being constrained by one that achieves excellent results while keeping the complexity low and a great capability of generalization for new samples. This means that for every new sample -product- the industry introduces to its production line, the training time for it should be minimum in order to provide an optimal solution. At the same time, since it would be impossible to predict all types of potential intrusions, it is important to verify the generalization capabilities of the implemented system, such that it’s sufficiently robust to correctly label samples non-existent in the training set. On top of that, the classification gets analyzed in terms of the number of false negatives it produces. The requirement is for zero of them, since otherwise the industry would be sending a potentially harmful product without them knowing. The second goal of this research proves a much harder task than the previous one. It encompasses the same constraints as before, while also having to distinguish between the different materials. By doing so, it can help the industry to identify the source of contamination, thus allowing quality managers to foresee potential faults in production lines and consequently prevent possible intrusions in packaged products. The results obtained achieved exceptionally accurate results for the binary part. Numerous ML models showed stable convergence and superb generalization capabilites. Regarding multi-class classification, it showed promising performance given the lower number of samples for each category.