A Pseudosupervised Approach for Food Spoilage Detection Using Metal Oxide Semiconductor Gas Sensors and Temperature Modulation

Food degradation leads to the emission of volatile organic compounds that accumulate in the headspace of sealed food containers and can be detected using metal oxide semiconductor (MOS) gas sensors. However, reliable labels for spoilage stages are often unavailable, and laboratory-based assays are destructive and time-consuming, limiting the use of conventional supervised learning approaches. To address this challenge, we propose a pseudosupervised framework that leverages signal patterns in sensor responses to estimate spoilage progression without external labels. Two independent strawberry spoilage measurement campaigns were conducted using commercially available MOS sensors operated under temperature-cycled operation (TCO). The evolving headspace composition during spoilage was continuously monitored in sealed containers. Data from five containers were used for model development, while two independent containers from a fresh strawberry batch were used for cross-batch validation. The proposed framework transforms monotonic sensor responses into temporally ordered pseudolabels, with transition buffering to ensure stable stage boundaries. Samples far from transition regions (i.e., excluding boundary-adjacent samples) were used to train an adaptive boosting (AdaBoost) classifier with depth-limited decision trees. The trained model achieved accuracies of 84.60% and 76.71% on two unseen containers, with weighted F1-scores of 83.5% and 75.5%, respectively. These results demonstrate that spoilage-stage classifiers can be derived directly from MOS-TCO responses without external chemical labeling, enabling noninvasive and data-driven food monitoring under cross-batch variability.