Deep Feedforward Neural Network-Based Soft Sensor for Industrial Process Modeling

Abstract

Soft sensors play a vital role in modern industrial processes by enabling real time estimation of critical process variables that are either difficult, expensive, or even impossible to measure directly using physical sensors. This study presents the application of Deep Feedforward Neural Networks (DFNNs) for soft sensor modeling in highly complex, dynamic, and non-linear industrial environments. DFNNs are designed and trained to accurately capture intricate and often hidden relationships between readily available input process measurements such as flow rates, temperatures, and pressures and essential output variables, including product quality indicators and performance metrics. The approach is validated using the Tennessee Eastman Process (TEP), a widely recognized and extensively studied benchmark for testing advanced process monitoring and control techniques. Simulation results demonstrate that the DFNN based soft sensor offers high prediction accuracy, strong generalization capabilities, and robust performance even under varying and uncertain operating conditions. These promising results confirm the potential of DFNNs as a powerful tool for soft sensing in a wide range of industrial applications.