Deep Learning based Prediction of Clogging Occurrences during Lignocellulosic Biomass Feeding in Screw Conveyors

Over the last decades, there have been substantial government and private sector investments to establish a commercial biorefining industry that uses lignocellulosic biomass as feedstock to produce fuels, chemicals, and other products. However, several biorefining plants experienced material conveyance problems due to the variability and complexity of the biomass feedstock. While the problems were reported in most conveyance unit operations in the biorefining plants, screw conveyors merit special attention because they are the most common conveyors used in biomass conveyance and typically function as the last conveyance unit connected to the conversion reactors. Thus, their operating status affects the plant production rate. Therefore, detecting emerging clogging events and, ultimately, proactively adjusting operating conditions to avoid downtime is crucial to improving overall plant economics. One promising solution is the development of sensor systems to detect clogging to support automated decision-making and process control. In this study, two deep learning based algorithms are developed to detect an imminent clogging event based on the current signature and vibration signals extracted from the sensors connected to the benchtop screw conveyor system. The study focuses on three biomass materials (switchgrass, loblolly pine, and hybrid poplar) and is designed around three research objectives. The first research objective examines the relationship between the occurrence of clogging in a screw conveyor and the current and vibration signals on the different feedstocks to establish the presence of clogging event fingerprint that could be exploited in automated decision-making and process-control. The second research objective applies two deep learning algorithms to the current and vibration signals to detect the imminent occurrence of clogging and its severity for decision making with an optimization procedure. The third objective examines the robustness of the optimized deep learning algorithm to detection imminent clogging events when feedstock properties (size distribution and moisture contents) vary. In the long-term, the early clogging detection methodology developed in this study could be leveraged to develop smart process controls for biomass conveyance.