A Multi-Step Sequence-to-Sequence Model with Attention LSTM Neural Networks for Industrial Soft Sensor Application

Abstract

Soft sensor technology is widely applied in the process industries to handle highly nonlinear, dynamic, time-dependent sequence data of industrial processes for predicting the key quality variables associated with auxiliary process variables. Many existing deep-learning-based soft sensor algorithms can build complex nonlinear models while rarely considering the dynamic characteristics of industrial data. In this paper, a multi-step sequence-to-sequence model with attention LSTM neural networks that is good at solving problems highly related to strong time-varying features is proposed to apply to the soft sensor modeling of industrial processes. The LSTM-based encoder-decoder structure with the attention mechanism layer is used to capture inherent features relevant to the quality variables and extract the long-term dependence of the sequential data. Instead of the normal sequence-to-point architecture, the improved sequence-to-sequence architecture predicts the multi-step quality variables, and the intermediate results are fed to a one-dimensional convolution module to get the final prediction results. By conducting industrial debutanizer column experiments, the superior performance of the model are demonstrated.