Dynamic Features Incorporated Locally Weighted Deep Learning Model for Soft Sensor Development

Abstract

Local learning models have been widely applied for time-variant process soft sensor development, where historical samples that are similar to the online testing sample are selected for local modeling and prediction. However, those similarity measurements and local models are commonly established based on the static samples, which are seriously inadequate when process dynamics exist. To tackle this problem, a novel locally weighted deep learning algorithm that takes dynamic features to describe both the similarities and relationships is proposed in this article. First, a deep dynamic feature extracting (DFE) network is constructed based on the long short-term memory (LSTM) encoder–decoder with attention mechanism, which maps time-sequence samples to a group of hidden dynamic features. With the extracted features incorporated into the original input features, a locally weighted autoencoder regression (LWAER) network is proposed for soft sensor modeling. Meanwhile, since both networks consist of unsupervised feature extracting and supervised feature regression, a large scale of unlabeled samples can be utilized in the semisupervised learning of model parameters. Finally, the superiorities of the proposed dynamic features incorporated LWAER (DFI-LWAER) model is verified by the outstanding performances in two real industrial soft sensing cases.