Nowadays, food companies produce large volumes of packaged food products to satisfy the population’s demand, which keeps increasing. Maintaining consumer trust and brand reputation is essential for such companies. However, packaged food products often conceal various contaminants that can compromise product quality, impact consumer health, and result in product recalls. Marler Clark, an American food safety law firm, gathers all incidents where products have been recalled because of undesired contaminants. According to the firm’s data, from 2018 to 2023, 199 recalls were made. In Italy, the Ministry of Health, responsible for product recalls, revoked 157 products because of food contamination. Insights from such entities underscore the frequent occurrence of product recalls resulting from contaminants like plastic and harmful bacteria. Contaminant presence harms brand reputation and customer loyalty, leading to substantial economic losses through compensations, penalties, and discarded production batches.
A company must prevent all physical, chemical, and biological contaminants from ending in a finished product. As food production quality assurance advances, destructive and non-destructive inspection controls play crucial roles in this task. The former requires measurements directly on the products (such as their temperature or oil content), while non-destructive techniques use faster and more efficient technologies like infrared. These approaches encounter notable limitations, particularly in detecting contaminants with low-density compositions or those concealed within metal packages. Our original research team proposed an innovative solution combining Microwave sensors and Machine Learning (ML) algorithms to address these challenges.
This novel methodology exploits the dielectric contrast between the inspected material and potential intrusions.
Through electromagnetic waves operating at specific frequencies, the presence of a foreign object disrupts wave patterns, enabling the detection of contaminated products via ML algorithms.
Building upon the pioneering work, the team currently working on the project aims to advance the existing solution in five directions.
To begin, we enhance the prototype pipeline used for the experiments. The current setup heavily relies on a specific component, a Vector Network Analyzer (VNA), which represents the whole system’s main cost and introduces several maintenance issues. The proposed alternative setup utilizes cheaper components while guaranteeing the same efficiency as the VNA.
The second step involves expanding the current dataset upon which the existing solution is built. This dataset expansion aims to strengthen further the tests we conduct on the robustness of the solution and, from a training perspective, enhances the algorithm’sability to identify contaminants with fewer samples accurately. The goal is to have a model able to adapt to different products with very different behavior regarding their “dielectric properties”. Therefore, we select a wide range of products to inspect: a carbonated and a low-CO2 soft drinks, soy sauce, flour, and honey. The chosen contaminants are plastic, paper, wood, glass, aluminum, glue, and cork, the most common physical contaminants that may fall into the product during production. Samples are also collected to cover different positions where the contaminant can realistically be found.
Thirdly, a theoretical examination of how most relevant biological contaminants impact food is performed.
Drawing insights from the most frequently recalled food products by the FDA, we identify prevalent contaminants and evaluate the latest techniques in managing them. Biological hazards produce severe risks to human health if consumed. Since viruses and parasites generally cause illness, most risks can be reduced by following Good Manufacturing Practices. Even so, we found that some bacteria, Salmonella spp.; Listeria monocytogenes; and Escherichia coli are still a persistent problem in food manufacturing. Here, the water activity can be a crucial parameter to help determine whether a product is contaminated based on the metabolic activities. Water activity levels above 80% are conducive to microbial growth. When the water activity is higher than 0.85, heat treatment is required. The next logical step in the evolution of our technology is to focus on products with high water activity levels (such as juice, milk and butter).
Additionally, we comprehensively evaluate state-of-the-art machine learning and deep learning models compared to the initial solution proposed by Wavision (based on a neural network). Performing some robustness tests on the initial solution, we notice that it suffers from a number of shortcomings. In particular, it fails to attain comparable levels of accuracy on different types of products and contaminants. To improve the detection accuracy while minimizing the calibration time for a novel implementation and further supporting real-time processing, we adopted an ensemble of the most effective methods identified through our analysis (Lasso and AdaBoost). Given that a company’s primary goal should be preventing the release of contaminated products into the market, thus minimizing false negatives, we focused on maximizing accuracy and achieving a zero false-negative rate. Results demonstrate promising outcomes, showcasing the potential for improved accuracy and real-time detection feasibility.
Finally, we introduce a Graph Neural Network (GNN) model trained to determine whether a container is contaminated, the type of contaminant, and a rough estimation of its position simultaneously. Given its higher parameter count and consequentially high inference time, such a model is devised to be used in a post-detection phase. The GNN architecture allows the scattering matrix to be naturally interpreted as a set of adjacency matrices by encoding the signal’s features onto nodes and edges of the graph. We train the model to perform classification over both contaminant type and position in the medium, obtaining near-perfect results in tests spanning various contaminants and media. Such a technique provides the company with a powerful tool to locate and address faults in the industrial production chain systematically.
While progress has been made, some research questions remain open. Despite the system works well in the laboratory prototype, its real-world implementation requires addressing new challenges such as limited computational resources and diverse environmental conditions.
Additionally, the study’s scope of contaminants and food products may not fully reflect what food companies encounter, which requires further research and validation along with different packaging materials. Despite these limitations, the Wavision team’s innovative approach addresses many of the limitations in non-invasive contaminant detection methods today, aiming to help food companies save on costs, reduce waste, and ensure customer safety.