On Grain Security by Temperature Interpolation: A Deep Learning Method for Comprehensive Data Fusion in Smart Granaries

As an indicator of grain safety, grain temperature data assumes great importance in the analysis of grain storage conditions and the decision-making of preventive measures such as ventilation and cooling. However, obtaining a thorough picture of grain temperature distribution via grain IoT with sensors deployed in the granary remains a challenge, given numerous data gaps across various areas due to insufficient coverage of the sensor network that fails to encompass the entire granary. Interpolation of grain temperature data, in this regard, is able to fill in the "unsensored" areas that are vacant in the records of data. Yet little literature is found in the frontier scholarship of grain temperature interpolation. To fill this noticeable niche, this study develops a novel data fusion interpolation model named convolutional neural network-attention-multilayer perceptron neural network (CAMNN) featuring an integration of convolutional neural network (CNN), attention mechanism, and multilayer perceptron (MLP). CNN is used to capture local spatial features of the temperature data, the attention mechanism enables the location of key and sensitive temperature areas, and MLP is incorporated for deep feature fusion. Performances of the proposed model are evaluated in a bin granary located in Shaanxi, China, and further validated in a larger bin granary of different storage types situated in Ningxia, China. Comparative assessments are conducted with five machine learning and deep learning (DL) models. Results indicate that CAMNN outperforms the other models, with a mean absolute error (MAE) of 0.5251 and a mean square error (mse) of 1.0881, demonstrating robust cross-context applicability across bin granaries varying in terms of sizes, storage types, and climatic zones.