Neural network models for predicting perishable food temperatures along the supply chain

Monitoring the temperature of perishable food along the supply chain using a limited number of temperature sensors per shipment is required for wide-scale implementation of quality-driven distribution. In this work, we propose to leverage the theoretical foundation and generalisation ability of a physical heat transfer model to develop a flexible neural net framework which can predict temperatures in real-time. More specifically, the temperature distribution inside a pallet subjected to different ambient temperatures are generated from a validated heat transfer model, and used to train a neural network. Simulations show that the neural network can predict the temperature distribution inside a pallet with an average error below 0.5 K in a one-sensor-per-pallet scenario when the sensor is properly located inside the pallet. Placing the temperature sensor at the corner of the pallet provides a high information content with strong correlations to the other locations inside the pallet to maximise the accuracy of the temperature estimates. The application of an ensemble operator to combine the predictions from multiple randomly seeded neural networks improved by up to 35% the accuracy of the temperature estimates. Finally, the introduction of small Gaussian noise in the training data is an efficient approach to improve the generalisation ability of the neural network and improved by nearly 45% the accuracy of the temperature prediction in the presence of noisy temperature sensors.