Abstract: To solve the difficulties caused by complex coupling between variables and obvious sequential characteristics in process production,a quality prediction method in multi-process production was proposed by combining multi-channel CNN-BiGRU and Temporal Pattern Attention(TPA).A gated convolutional network composed of Convolutional Neural Network(CNN) and Bidirectional Gate Recurrent Unit(BiGRU) was built to obtain the nonlinear time dynamic correlation of process data in multi-process production,and the high-dimensional feature vectors reflecting the sequential variation of process parameters were constructed into time series,which were input into the forward and backward GRU networks respectively to avoid the problem of gradient disappearance or explosion when training the long sequential process data;the TPA was introduced to adaptively allocate attention weights for different process state variables in the production process,hence obtained the associated coupling characteristics between different process parameters,and the final quality prediction through the fully connected layer;the quality prediction experiment was carried out using the data set of five processes of a tobacco production line.The Mean Absolute Error(MAE) and Root Mean Square Error(RMSE) were reduced by more than 21.36% and 26.56% compared with TCN-Attention and DA_BiLSTM models.Results showed that the CNN-BiGRU-TPA model effectively improved the prediction accuracy,which provided an available method for quality prediction of multi-process production.

Key words: process manufacturing, time series feature prediction, temporal pattern attention mechanism, multi-process coupling

摘要： 针对流程生产由于变量间耦合复杂、时序特征显著而导致工艺质量精准预测困难问题,提出了一种融合多通道CNN-BiGRU与时间模式注意力机制的多工序工艺质量预测方法。首先,搭建由卷积神经网络(CNN)与双向门控循环单元(BiGRU)组成的门控卷积网络,用于获取多工序生产过程工艺数据的非线性时间动态相关性,并将反映工艺参数时序变化规律的高维特征向量构成时间序列,分别输入到前向和后向传递的GRU网络,避免在训练工艺数据的长时间序列时的梯度消失或梯度爆炸问题;其次,引入时间模式注意力机制(TPA)为生产过程中的不同工序状态变量自适应分配注意力权重,动态获取不同工艺参数之间的关联耦合特征,通过全连接层获取最终工艺质量的预测结果;最后,利用某制丝产线五大工序的工艺数据集进行了工艺质量的预测实验。实验表明,相较于TCN-Attention和DA_BiLSTM等模型,CNN-BiGRU-TPA模型有效提高了预测精度,平均绝对误差(MAE)和均方根误差(RMSE)降低了21.36％和26.56％以上,为流程生产多工序质量精准预测提供了实现方法和途径。

关键词: 流程制造, 时序特征预测, 时间模式注意力机制, 多工序耦合