Abstract: Due to the serious coupling of energy,mass flow between process manufacturing processes and significant time series characteristics of process parameters,the accurate prediction of product quality has become an urgent problem to be solved under the existing manufacturing mode.Based on analyzing the characteristics of multi-dimension,strong time sequence and correlation coupling of process manufacturing process performance index,an Attention Model—Gated Recurrent Unit—Back Propagation Neural Network (AM-GRU-BPNN) based multi-process coupling parameter correlation prediction method was proposed.The polymorphic and heterogeneous production data were screened by mutual information method as input to establish a ConvGRU autoencoder.The temporal features of process data,process parameters,operating parameters and other data through unsupervised learning were extracted.At the same time,attention mechanism was introduced to extract the coupling correlation features of different processes and embed the vectors,so as to allocate attention weights to the process parameters of different processes.On this basis,the attention network self-learning was designed to study the influence of process correlation characteristics on quality performance indicators at different times.Then,the important correlation features were enhanced through GRU network,and the single process prediction model was aggregated according to the time sequence characteristics to realize the time sequence coupling feature fusion of multiple processes.The accurate prediction of product process quality was completed through the output layer BPNN.Experimental results showed that AM-GRU-BPNN could effectively improve the prediction accuracy and provide the basis for the process control of production line from the angle of multi-procedure.

Key words: process manufacturing, multi-process coupling, attention model—gated recurrent unit—back propagation neural network, temporal sequence feature fusion, correlation prediction

摘要： 鉴于流程制造工序间能质流耦合严重,性能指标影响因素众多,工艺参数时序特征显著,现有制造模式下难以精准预测产品质量,在分析流程制造工艺性能指标多维、强时序、关联耦合特征的基础上,提出一种基于注意力机制—门控循环单元-BP神经网络(Attention AM-GRU-BPNN)的多工序耦合参数关联预测方法。首先采用互信息方法筛选多态异构生产数据作为输入,建立ConvGRU自编码器,通过无监督学习对过程数据、工艺参数、操作参数等进行时序特征提取,同时引入时序注意力机制提取不同工序的耦合关联特征并进行向量嵌入,为不同工序的工艺参数分配注意力权重。在此基础上,设计Attention网络自学习不同时刻下工艺关联特征对质量性能指标的影响差异,再通过门控循环单元网络对重要的关联特征进行增强,并按照时序特征对单工序预测模型进行聚合,实现多工序时序特征融合,最后通过输出层BPNN神经网络精准预测产品工艺质量。实验表明,AM-GRU-BPNN有效提高了预测精度,从多工序角度为生产线工序的加工过程控制提供了依据。

关键词: 流程制造, 多工序耦合, 注意力机制—门控循环单元-BP神经网络, 时序特征融合, 关联预测