A Model Combining Seq2Seq Network and LightGBM Algorithm for Industrial Soft Sensor

Abstract As a key technology for industry 4.0, data-driven soft sensing plays an important role in the control and optimization of industrial processes. However, due to the large-scale, nonlinear and dynamic characteristics of industrial data, it is difficult to process industrial data. To solve these difficulties, a soft sensor modeling method based on a sequence to sequence model and a gradient boosting tree algorithm is developed. In this method, an unsupervised trained Seq2Seq model is used to extract dynamic features at first. Then a high-precision model based on LightGBM is constructed with the dynamic features and the original features as inputs. The developed method is validated on pulping data and compared with other machine learning methods such as RNN and SVR. The result shows the developed method has a better performance.