Deep learning-based ranging error mitigation method for UWB localization system in greenhouse

Abstract

The intelligent devices in indoor agriculture put forward high requirements for localization. However, the narrow space and various obstacles in the greenhouse generate non-line-of-sight (NLOS) conditions for wireless signal propagation, which challenge the quality of localization. To solve this problem, we proposed a ranging error mitigation method using deep learning, based on the characteristic that channel impulse response (CIR) presents different response states under different NLOS conditions. The method took the CIR as input, trying to predict the ranging error of wireless module and then mitigated the positioning results. To obtain useful features from massive data, this paper adopted the state-of-art residual network model ECA-ResNet with a lightweight attention mechanism module as the deep learning framework. All the operations were executed in the ultra-wideband (UWB) localization system with five anti-interference and precise DWM1000 wireless transceivers. Through the training of the collected range dataset, the model performed well in a complex environment with MAE, RMSE and R2 of 0.1004 m, 0.1714 m and 0.7548. Integrated with the above model, the UWB localization system was tested for positioning in a greenhouse with multiple compartments. It showed that large estimated positioning errors were significantly mitigated, which maintained the usability of results in such a heavily obstructed environment.