Development of a NARX neural network for a tribo-aero-electrostatic separator with rotating disk electrodes

• Datum: 2026
• Hauptverfasser: Ayache, Z., Dahou, O.
• Format: Artikel
• Sprache: Englisch
• Veröffentlicht: National Technical University "Kharkiv Polytechnic Institute" and Аnatolii Pidhornyi Institute of Power Machines and Systems of NAS of Ukraine 2026
• Schlagworte: electrostatic separation; high voltage; dynamic modeling; NARX neural network; recycling
• Online Zugang: http://eie.khpi.edu.ua/article/view/339775
• Назва журналу: Electrical Engineering & Electromechanics

Introduction. The exponential growth of waste electrical and electronic equipment (WEEE) requires efficient strategies for plastic waste management. Plastics, a major fraction of WEEE, represent both an environmental challenge due to low biodegradability and a valuable source of secondary raw materials. Problem. Tribo-aero-electrostatic separators with rotating disk electrodes offer a promising solution for fine plastic separation. However, their performance depends on multiple, nonlinear, and time-varying factors such as disk speed, voltage, and particle properties. These complex interactions make analytical modeling and stable process control difficult, limiting industrial implementation. The goal of this work is to develop a reliable dynamic model based on NARX neural networks capable of predicting the real-time evolution of key process variables such as recovered mass and particle charge. Methodology. The proposed NARX neural network learns temporal nonlinear relationships directly from experimental data, avoiding the need for explicit physical equations. Experiments were conducted on a synthetic 50:50 mixture of Acrylonitrile Butadiene Styrene (ABS) and Polystyrene (PS) particles (500-1000 μm) to assess model performance under varying disk speeds, voltages, and air flow rates. Results. The developed model accurately predicts the recovered mass and acquired charge of both ABS and PS over a wide range of operating conditions. The predictions show strong agreement with experimental measurements, maintaining low error levels even at parameter extremes. Scientific novelty. This work represents the first application of NARX neural networks to model the dynamic behavior of a two-rotating-disk tribo-aero-electrostatic separator. The approach captures essential time-dependent interactions that conventional static or analytical models fail to describe. Practical value. The NARX model exhibits high predictive accuracy and robustness across an extended operating domain (4–20 kV, 15–60 rpm, 7–9 m3/h), with errors limited to the 10–3 g and 10–3 µC ranges. These characteristics demonstrate its potential for real-time intelligent control and adaptive optimization of electrostatic separation processes in plastic waste recycling. References 39, tables 3, figures 9.