研究“三分光束”光内同轴送丝激光熔覆各工艺参数的工艺区间及参数与熔覆层几何形貌映射关系。首先，采用单因素实验方法研究激光功率、扫描速度、送丝速度、离焦量四个工艺参数的工艺区间；其次，以熔覆层的高度、宽度、横截面积作为熔覆层几何形貌的量化指标；最后，分别建立神经网络模型和二次回归模型实现熔覆工艺参数和熔覆层形貌量化指标之间映射关系的预测。基于单道单因素实验，当激光功率介于1 300~1 700 W，扫描速度介于3~7 mm/s，送丝速度介于9~15 mm/s，离焦量介于?2.5~?1.5 mm时能获得液桥过渡熔覆质量较好的单道。在对测试样本数据的预测中，在置信度85%情况下，BP神经网络模型对熔覆层高度、宽度、横截面积的预测精度分别为100%, 100%, 93.33%，均方根误差分别为0.21, 0.07, 0.24；二次回归模型的精度分别为100%, 66.67%, 73.33%，均方根误差分别为0.21, 0.13, 0.28。结论说明二次回归模型中变量的交叉项未能拟合送丝熔覆多变量耦合的非线性过程，而BP神经网络得到较好的预测结果。

The work aim to study parameters window of “three beam” coaxial wire feeding and the mapping relationship between parameters and cladding geometry. Firstly,the process interval of four process parameters of laser power, scanning speed, wire feeding speed and defocusing amount was studied by single factor experiment method; Secondly, the height, width and cross-sectional area of the cladding layer was used as the quantitative indicators of the geometry of the cladding layer; Finally, a neural network model and the quadratic regression model were set up respectively which were used to predict the mapping relationship between the cladding process parameters and the quantitative indicators of the cladding layer. Based on single-channel single-factor experiments, when the laser power was between 1 300 W and 1 700 W, the scanning speed was between 3 mm/s and 7 mm/s, the wire feeding speed was between 9 mm/s and 15 mm/s, and the defocusing amount was between ?2.5 mm and ?1.5 mm can get the cladding of liquid bridge transition with good quality. Besides, in the prediction of the test sample data, under the condition of 85% confidence, the prediction accuracy of the BP neural network model for the height, width and cross-sectional area of the cladding layer is 100%, 100%, 93.33%, and the root mean square error is 0.21,0.07,0.24. The accuracy of the quadratic regression model is 100%, 66.67%, and 73.33%, respectively, and the root mean square errors are 0.21, 0.13, and 0.28, respectively. From the result, the cross terms of the variables in the quadratic regression model failed to fit the nonlinear process of wire cladding. By contrast, BP neural network obtained better prediction results.