基于时空全卷积循环神经网络的零件表面形貌预测

A Spatio-temporal Fully Convolutional Recurrent Neural Network Based Surface Topography Prediction

零件表面形貌预测对于降低产品质量波动和加工成本、减少零件废品率有重要意义。基于高清晰测量数据,提出一种时空全卷积循环神经网络非平稳时空序列预测模型,实现零件加工表面三维形貌的预测。通过计算全局莫兰指数和时间自相关函数进行时空相关性分析,为模型构建准确的输入,并克服传统预测方法未充分利用数据全局特征和局部特征的缺点以及模型输入的随意性。实例研究的结果表明,提出的方法具有更优的综合预测效果,其预测精度优于传统方法 12%~18%,预测时间是传统方法的1/5。

Surface topography prediction is of great significance to reduce the fluctuation of product quality and processing cost as well as the scrap rate of parts. Based on the high definition metrology(HDM) measured data, a spatio-temporal series prediction model called spatio-temporal fully convolutional recurrent neural network(STFCRNN) is proposed, which achieves the 3 D surface topography prediction of machining surface. The time autocorrelation function and global Moran’s I are calculated to analyze the spatiotemporal correlation of the surface data, which is used to conduct an accurate input of the model, thereby overcoming the shortcomings of traditional prediction methods that do not make full use of the global and local features of the data and the randomness of model input. Moreover, the result of case study shows that the proposed method has a better comprehensive prediction performance, and its prediction accuracy is 12%-18% better than that of traditional methods and the prediction time is 1/5 of traditional methods.