Enhanced Machine-Learning Flow for Microwave-Sensing Systems for Contaminant Detection in Food

Combining data-driven machine learning (ML) with microwave sensing (MWS) makes it possible to analyze packaged food in real time without any contact and spot low-density contaminants (e.g., plastics or glass splinters) that current industrial food safety methods cannot detect. This is achieved by training ML classifiers on the scattered signal reflected by the target food product exposed to MWs. In this article, we present an enhanced ML flow to boost foreign body detection accuracy of ML classifiers in MWS systems. Innovations include assessing the performance of a multiclass classifier, training it with MW frequency pairs as features, data augmentation, a new preprocessing scaler suitable for the feature distributions in the datasets, quantization, and pruning. The final test results, obtained using our previously designed MWS system and collected dataset of contaminated hazelnut-cocoa spread jars, show a multiclass accuracy for the floating-point model of 96.5%, which corresponds to a binary-equivalent accuracy of 97.3%. This is an improvement of +3.3% against the binary classifier of the previous work. The quantized and pruned model, instead, reached a multiclass accuracy of 94.2%, or a binary accuracy of 95.4%, thus still improving the previous work by +1.4%. Also, we achieved a latency of 26 $\mu$s on an AMD/Xilinx Kria K26 field programmable gate array (FPGA), a result which is ideal for high-throughput food production lines. Furthermore, we expand our latest work with supplementary details and experiments to further validate the proposed ML flow, including a comparative analysis against our prior results. Lastly, we share our datasets publicly on OpenML.