Improved Bi-LSTM With Distributed Nonlinear Extensions and Parallel Inputs for Soft Sensing

Industrial soft sensing models have found extensive application in predicting key process variables that are challenging to directly measure. However, the effectiveness of conventional soft sensing models is impacted by the intricate characteristics of process variables, such as high nonlinearity, coupling, and complex dynamicity. To address this limitation, an enhanced bidirectional long short-term memory (Bi-LSTM) model based on distributed nonlinear extensions integrated with parallel inputs (DNEPI-Bi-LSTM) is proposed for constructing the soft sensing model. First, to account for the differential impact between inputs and outputs, partial correlation is employed to segregate the inputs into two categories: positive subinputs and negative subinputs. Subsequently, these two distributed subinputs are transformed into nonlinear space by passing through the hidden layer of the extreme learning machine. The resulting outputs from the hidden layer are considered as distributed nonlinear extensions. Finally, the enhanced DNEPI-Bi-LSTM soft sensing model is developed using parallel inputs integrated with distributed nonlinear extensions. To assess the efficacy of DNEPI-Bi-LSTM, an industrial process known as the sulfur recovery unit is adopted. Simulation results illustrate that DNEPI-Bi-LSTM outperforms other advanced models in terms of accuracy, showcasing its potential in industrial applications.