Development of a virtual sensor for temperature-break detection: A novel ensemble input-variable-selection method

This study utilises a systematic approach for design and development of an efficient One-Sensor-Per-Pallet Virtual Temperature Sensor (OSPP-VTS) for cold-chains. Temperature monitoring in such facilities is desirable with maximum spatial resolution possible, while utilising the minimum number of physical sensors. To realise this objective, a novel ensemble supervised Input-Variable-Selection (IVS) method was introduced to identify the most suitable source-points for the time-delay neural network, as the estimator algorithm. The ranking derived from this IVS method was then used to reduce the number of the required source-points. The reliability of the ranking was also investigated using meta-heuristic methods. Out-of-Sample evaluation of the OSPP-VTS demonstrated that the system is capable of accurately estimating temperature at twenty locations of the fruit pallet without any prior knowledge about its condition. This resulted in RMSE_mean of 0.67 K which is Satisfactory when extensive data acquisition under various temperature scenarios is not an option. The outcomes indicated that the novel IVS method shows promise as a non-subjective systematic approach for determining the best locations and ranking them, if necessary, with its ranking 100 % matching the Differential Evolution Algorithm. The developed system exhibits the capability of detecting the occurrence of temperature breaks. Evaluation results demonstrated that this approach can be leveraged to improve the spatial resolution of cold chain temperature monitoring by providing a robust framework for virtual sensor development.