Machine-Learning Classification of Solids Mass Flow Rate in Gas–Solid Pneumatic Conveying Using Electrostatic Sensor Measurements

Electrostatic signals in pneumatic conveying depend on solids mass flow rate, air velocity, particle distribution and material, complicating direct estimation. This study classified five nominal solids mass flow rates (0.1–10 g/s) for wheat flour and ground coffee at four air velocities (15–30 m/s) using six ring electrodes and seven arc segments with guarded electrometer-grade front ends. From 1 s blocks, 29 time-, frequency- and time–frequency-domain features per channel were extracted. The Common Vector Approach (CVA), support vector machines (SVMs), multilayer perceptrons (MLPs), random forests (RFs) and feature-based convolutional neural networks (CNNs) were compared with an end-to-end CNN on raw signals. With one feeding cycle per record held out, the best balanced accuracies were 86.9% (flour) and 89.6% (coffee), and 97.8–100% of predictions were within one class. RF regression yielded typical error factors of 1.34 (flour) and 1.31 (coffee). Depending on the classifier, balanced accuracy declined to 47–70% for unseen air velocities and 26–59% across materials; a no-flow control suggested possible contributions from record-specific offsets to the accuracy. These findings restrict the reported performance to the tested conditions and support velocity-range coverage, material-specific calibration and recording-order controls.