Deep transfer model with source domain segmentation for polyester esterification processes

In the polyester fiber industry, newly-built production lines or newly-installed production equipment cannot collect large enough historical data, resulting in the inability to construct accurate time series prediction models for key indicators. This paper proposed a deep transfer model with source domain segmentation (SDS-DTM), which transfers knowledge of another production line with sufficient time series data, and the proposed model is used for multivariate time series forecasting modeling in target production line with insufficient data. SDS-DTM adds a domain adaptation layer to the deep recurrent neural network, which transfers knowledge by minimizing distribution mismatch in the feature sharing space. The long time span of the source time series will cause the problem of temporal covariate shift (TCS). SDS-DTM minimizes the influence of TCS by splitting the source domain and constructing an integrated model. The effectiveness of the proposed model is verified by polyester esterification process data.