Abstract:  To address the problems which were difficult to accurately predict production quality due to the temporal coupling of multiple processing parameters in process production, a fast and efficient production quality prediction method was proposed based on deep TCN networks and migration learning. With a sequence-to-sequence learning structure, a deep TCN and a temporal attention mechanism formed the encoding component for extracting key temporal features from multiple sources, while a residual long short term memory network formed the decoding component for simultaneous extraction of quality temporal information, and migration learning was introduced to address the adaptability of the prediction model to online production quality prediction. The experiments show that the proposed method has significant advantages in prediction accuracy and stability, and has high prediction accuracy and computational efficiency in predicting small sample data.

Key words: process quality, time series correlation prediction, sequence to sequence, temporal convolution network(TCN), transfer learning

摘要： 针对流程生产多工艺参数时序耦合导致的生产质量难以准确预测的问题，提出了基于深度时间卷积神经网络与迁移学习的生产质量快速高效预测方法。借助序列到序列的学习框架，采用深度时间卷积神经网络和时序注意力机制构成的编码器提取多源关键时序特征，采用残差长短期记忆神经网络构成的解码器挖掘质量时序信息，引入迁移学习解决预测模型对生产质量在线预测适应性的问题。实验表明所提方法的预测精度与稳定性优势显著，且在小样本数据预测时具有较高的预测精度和计算效率。

关键词: 工艺过程质量, 时序关联预测, 序列到序列, 时间卷积神经网络, 迁移学习