基于时序深度学习的数控机床运动精度预测方法

摘要:

由于数控机床精度演化规律难以通过数学建模分析，提出了一种基于时序深度学习网络的数控机床运动精度建模与预测方法。基于长短时记忆网络建立时序深度学习预测模型，利用相空间重构原理构建模型时序输入向量，采用多层网格搜索方法选择最优隐含层层数、隐含层节点数等模型参数，以BPTT方法训练模型；模型自动提取运动精度时间序列的时空特征，挖掘精度时间序列前后关联信息，对运动精度变化趋势进行预测。实验结果表明，基于时序深度学习网络的预测模型能够准确预测数控机床精度的衰退趋势，预测的最大相对误差不大于7.96%，优于传统方法。

Abstract:

Because of the difficult to analyze the evolution law of CNC machine tools accuracy through mathematical modeling, a method of motion accuracy modeling and prediction based on sequential deep learning network was proposed. A deep learning network was presented based on the long shortterm memory (LSTM). Using the principle of phase space reconstruction, the sequence input vector of the model was constructed. The optimal parameters of the model, such as number of hidden layer and number of hidden layer node were determined based on multilayer grid search algorithm. The model was trained with BPTT method. The mutual information before and after the precision time series was mined with data driven. The temporal and spatial characteristics of the motion accuracy series were automatically extracted through the deep learning network. Finally, the declining trend of motion accuracy was predicted by the model. The experiments results showed that the prediction model based on the sequential deep learning network could predict properly the evolutionary trends and regularity of the precision. The maximum relative error of prediction was not more than 796%. The prediction accuracy of the method was better than that of the traditional methods. The method was helpful for evaluating the reliability of NC machine tools and ensuring the machining accuracy.