Deep-learning-based Intelligent Navigation And Active Disturbance Rejection Control Of Weeders

In recent years,mechanical weeding,which causes no environmental pollution,has gradually become a crucial technology to promote the sustainable development of agriculture.However,the injury rate of rice can be beyond acceptable while using mechanical weeding,due to the complexity of the paddy fields.For a riding paddy field weeder,which is a mechanical weeding machine,the main reason for rice injury is the misalignment between the weeding wheels and the rice rows.In this paper,we study and design an autonomous correction system based on Deep Learning and Active Disturbance Rejection Control(ADRC)for the weeding wheels to solve the rice injury problem.The system consists of an image acquisition system and a hydraulic position servo system.The image acquisition system extracts the guidance line based on Deep Learning,and the hydraulic position servo system drives the weeding wheels to move laterally.Our main work in this paper is as follows:1.Rice detection and guidance line extraction algorithm based on Deep Learning:Images of rice seedlings about 15 days after transplanting are collected in the paddy fields,out of which we select 587 images and establish a novel rice seedling dataset.A Mobile Net V3-SSD model is trained on the dataset,which achieves 87.8% m AP on the test set at 69 FPS on an NVIDIA Ge Force GTX 1050.Guidance line information is extracted using the DBSCAN clustering algorithm and least-squares method based on the rice detection results,providing the position deviation for the hydraulic position servo system;2.Hydraulic position servo system based on ADRC: The mathematical model of the hydraulic system is analyzed and derived,based on which the transfer function and the open-loop gain are obtained through the system identification method.An ADRC controller is designed,and the parameters influence and system characteristics are studied through Simulink simulation.A nonlinear function modified from the Sigmoid function is proposed for the oscillation caused by the large-scale dead zone,with the residual dead zone compensated using the total disturbance estimation of the extended state observer(ESO).The method proposed above is proved to be effective through Simulink simulation;3.Implementation and experiments of the autonomous correction system: A three-stage experiment is designed.The first two stages are designed independently for testing the performance of the hydraulic position servo system.The last stage is designed for testing the injury rate in the paddy fields of the weeder,equipped with the autonomous correction system.For the hydraulic position servo system,the results show that the setting time is under 1second and there is approximately no overshoot.The steady-state error is under 0.5 mm on the experimental platform,and 6 mm on the weeder.With the weeder equipped with the autonomous correction system,the injury rate is below 5%,solving the rice injury problem caused by the misalignment between the weeding wheels and the rice rows.