Abstract: To address the challenge of nonlineofsight (NLOS) error interference in the interrow positioning of agricultural inspection robots, this study proposes an EWMAoptimized BP neural network positioning method. The model enhances the standard BP neural network by integrating an exponentially weighted moving average (EWMA)algorithm, which compensates for the BP networks reduced accuracy in longdistance positioning. EWMA adopts a cubic function to design weighting coefficients, thereby improving the reliability of position estimation. To mitigate slow convergence and local optima in the BP neural network, momentum factors and adaptive learning rates were introduced. Additionally, input and output vectors with large magnitude differences werenormalized to enhance computational efficiency and prevent neuron saturation. Finally, the Chan algorithm, which integrates the TDOA localization model, was employed to obtain the optimal positioning values. Experimental analysis of interrow static and dynamic positioning in an agricultural greenhouse demonstrated that the proposed model achieved a static positioning error of 90%, which does not exceed 0.2cm. In dynamic positioning, the average estimation error along the Xaxis remained within 3 cm, while the Yaxis average estimation error remained within 5 cm. Under NLOS interference, the proposed method reduced the average positioning error by 90%, effectively filtering out NLOSinduced errors. These results confirm the models robustness and high accuracy in agricultural inspection robot positioning.

Key words: agricultural inspection robots, accurate positioning, BP neural network, non line of sight

摘要： 为解决农业巡检机器人行间定位过程中产生非视距（NLOS）误差干扰，提出一种EWMA—优化BP神经网络模型定位方法。在BP神经网络基础上，融合指数加权移动平均算法（EWMA）弥补其定位远距离处精度较低的不足，同时EWMA采用三次函数形式设计加权系数；为避免BP神经网络收敛速度慢、局部最优，引入动量因子和自适应学习速率进行改进BP神经网络，并通过归一化处理数量级差距很大的输入、输出向量，考虑到计算方便，降低神经元饱和度，最终融合TDOA定位模型的Chan算法获得最优值。农业大棚巡检机器人行间动静态定位试验分析表明，所提模型静态定位误差90%的情况下不超过0.2 cm，动态定位X轴方向估计误差均值为3 cm，Y轴方向估计误差均值为5 cm，NLOS因素干扰下平均定位误差降低90%，基本滤除NLOS因素干扰下的定位误差。

关键词: 农业巡检机器人, 精准定位, BP神经网络, 非视距