Abstract: This study addressed the challenges of vortex formation and dead air zones in traditional long-pressure short suction ventilation and dust removal systems. By leveraging the Coanda effect, the system's structure was optimized through the nesting of the exhaust and pressure ducts. This design enhanced airflow in the negative pressure duct, promoting adherence to the tunnel walls, reducing dust dispersion, and significantly lowering energy consumption. Simulations utilizing flow field analysis and discrete phase model (DPM) revealed an optimal pressure-extraction ratio of 2∶3. Under this ratio, results indicated that the optimized system reduced dust concentrations at the driver's position and downwind side by 5.56% and 55.41%, respectively, compared to traditional systems. With the structure and pressure-extraction ratio established, further improvements in dust removal efficiency were achievable through parameter regulation. Key parameters included the distance between the duct and the dust-producing surface, the distance from the duct's central axis to the ground, and the distance between the pressure and extraction ducts. Convolutional neural network (CNN) was employed for intelligent parameter control, enabling the identification of optimal parameters for varying initial dust concentrations. Experiments conducted on 45 parameter regulation schemes using a scaled-down experimental platform demonstrated that the CNN model outperformed BP neural networks in accuracy and stability for dust concentration predictions. When initial dust concentrations at the driver's position and downwind side ranged from 300 to 900 mg/m³, the optimized system achieved an average dust concentration reduction of 51.49% to 83.88%, thereby validating the effectiveness of parameter control.