Abstract: In mixed-model assembly job shops,the level of work-in-process inventory between two stages of processing and assembly can lead to production interruptions or incorrect/missing assembly.To achieve precise control,accurate prediction of inventory levels at different times is crucial.A Temporal Convolutional Network—Bidirectional Long Short Term Memory (TCN-BLSTM) model integrated with Channel and Temporal Attention(CATA) was proposed.In this model,a denoising autoencoder was used for feature dimensionality reduction,removing noise and redundant information.The transmission mechanisms of workpiece flow between processing machines and assembly stations was captured with TCN,embedding channel attention within TCN to deeply extract key features.The BLSTM simulated the bidirectional information flow at push-pull moments,while the Temporal Attention Network with multiple modules enhanced the relationships between features across all time points,identifying critical moments affecting inventory.Backpropagation was used to update the attention network parameters for precise prediction.Experimental results showed that the proposed CATA-TCN-BLSTM model effectively identifies key features and moments and significantly improves prediction accuracy.Meanwhile,a transfer of models under different production scenarios is realized with a prediction accuracy of over 98%.

Key words: work-in-process inventory prediction, temporal convolution network, channel attention, bidirectional long short-term memory network, temporal attention

摘要： 混装作业车间加工和装配两阶段间的在制品库存水平,可能直接导致装配生产中断或引起错装漏装,故需预测不同时刻库存水平,实现精准控制。提出了融合通道与时间注意力机制(CATA)的TCN-BLSTM模型。首先,利用去噪自动编码器进行特征降维,去除数据噪声和冗余信息;通过时间卷积网络(TCN)捕获工件流在加工机器装配工位间的传递机制,并将通道注意力嵌入时间卷积网络中,挖掘关键特征;通过双向长短时记忆网络(BLSTM)模拟推拉时刻双向信息流的传递,设计含多个模块的时间注意力网络,增强所有时刻特征间关联关系,挖掘影响库存的关键时刻;通过反向传播更新注意力网络参数,实现精准预测。实验结果表明所提出的CATA-TCN-BLSTM模型有效挖掘了关键特征和时刻,大幅提高了预测准确率;并实现了不同生产场景下的预测模型迁移,预测准确率达98%以上。

关键词: 在制品库存预测, 时间卷积网络, 通道注意力, 双向长短期记忆网络, 时间注意力