基于FDA-LSTM的冷轧过程多源异构时序数据处理及颤振预测

Multi-source heterogeneous time series data processing and chatter prediction based on the method of FDA-LSTM in cold rolling processes

冷连轧颤振诱发机理复杂多变,颤振问题的解决需要通过大数据驱动的信息挖掘对机理模型进行补充。该研究针对某冷连轧机现场采集的工艺参数及振动数据,通过函数型数据分析(functional data analysis,FDA)方法进行预处理,实现多源异构时序数据的频率协同;采用SelectKBest算法对影响颤振的多种工艺参数进行特征选择,筛选出与振动相关性较强的因素,构造样本空间;基于长短时记忆(long short-term memory,LSTM)神经网络建立振动能量值的预测模型,并与径向基函数(radial basis function,RBF)神经网络、循环神经网络(recurrent neural network,RNN)模型进行比较。结果表明,LSTM模型具有较高的预测精度,同时采用阈值法验证该模型能有效地预测颤振的发生。

The induced mechanism of cold tandem rolling chatter is complex and changeable, and in order to solve the chatter problem, it needs to supplement a mechanism model through big data-driven information mining. Based on the process parameters and vibration data collected on-site in cold tandem rolling mills, through preprocessing the data by the functional data analysis(FDA) method, the frequency coordination of multi-source heterogeneous time series data was achieved, thereby improving the accuracy of the sample expansion. The feature selection for multi kinds of process parameters that strongly affect the chatter was made by use of the SelectKBest algorithm to construct a sample space. A prediction model for vibration energy was established based on the long short-term memory(LSTM) neural network, and the results were compared with those of the radial basis function(RBF) neural network and the recurrent neural network(RNN) models. The results show the higher prediction accuracy of the LSTM model. And the threshold method was used to verify that the model can effectively predict the occurrence of chatter.