基于ASSFOA-GRNN模型的激光定向能量沉积温度预测

Temperature Prediction of Laser Directed Energy Deposition Based on ASSFOA-GRNN Model

摘要:

针对传统的激光定向能量沉积（LDED）温度预测方法精度低、耗时长、成本高等问题，提出了一种结合数值模拟的机器学习模型来预测LDED过程中的温度。建立了有限元（FE）热分析模型，通过原位监测实验验证了模型的准确性，并基于FE模拟获得了预测模型的基础数据库。利用广义回归神经网络（GRNN）进行温度预测。为了减少GRNN调参过程中对人为经验的依赖并提升模型预测性能，引入了改进的自适应步长果蝇优化算法（ASSFOA）。最后将ASSFOA-GRNN模型与BPNN模型、GRNN模型和FOA-GRNN模型的预测性能进行了比较，评价指标包括均方根误差（RMSE）、平均绝对误差（MAE）、决定系数（R²）、训练时间和预测时间。结果表明，ASSFOA-GRNN模型在RMSE、MAE和R²指标上均表现最佳，尽管其预测效率略低于FOA-GRNN模型，但预测精度显著优于其他对比模型。本研究方法可用于LDED工艺的温度预测，也为同类方法的应用提供了借鉴。

Abstract:

To address the issues of low accuracy, long time consumption, and high cost of the traditional temperature prediction methods for laser directed energy deposition (LDED), a machine learning model combined with numerical simulation was proposed to predict the temperature during LDED. A finite element (FE) thermal analysis model was established. The model's accuracy was verified through in-situ monitoring experiments, and a basic database for the predictive model was obtained based on FE simulations. Temperature prediction was performed using a generalized regression neural network (GRNN). To reduce dependence on human experience during GRNN parameter tuning and to enhance model prediction performance, an improved adaptive step-size fruit fly optimization algorithm (ASSFOA) was introduced. Finally, the prediction performance of ASSFOA-GRNN model was compared with that of back-propagation neural network model, GRNN model, and fruit fly optimization algorithm (FOA)-GRNN model. The evaluation metrics included the root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), training time, and prediction time. Results show that the ASSFOA-GRNN model exhibits optimal performance regarding RMSE, MAE, and R² indexes. Although its prediction efficiency is slightly lower than that of the FOA-GRNN model, its prediction accuracy is significantly better than that of the other models. This proposed method can be used for temperature prediction in LDED process and also provide a reference for similar methods.