Vision-assisted optimization of coagulation–flocculation parameters using a joint YOLOv11 and U-Net + + framework

This study investigates a defined coagulation–flocculation–sedimentation unit for variable-composition chemical wastewater rather than treating wastewater treatment as a single universal process. A joint YOLOv11 and U-Net + + perception model was developed to identify suspended particles, oil films, flocs, and foam and to quantify their spatial coverage before periodic adjustment of pH, flocculant dosage, and hydraulic retention time. YOLOv11 provides object categories, locations, and confidence scores, whereas U-Net + + supplies pixel-level masks; their outputs are fused into a 128-dimensional state representation and four physically interpretable visual variables. The optimization layer uses a constrained multi-objective particle swarm procedure to balance predicted COD and suspended-solids removal against normalized chemical consumption. On the batch-separated test set, the detector achieved an [email protected] of 91.5% and a small-object recall of 79.4%, while the segmentation network achieved an IoU of 89.7%. Five repeated runs produced an [email protected]:0.95 of 0.884 ± 0.006 and a mean IoU of 0.897 ± 0.005. The complete image-to-decision cycle required 12.25 s, which is compatible with the five-minute sampling interval but is therefore described as near-real-time periodic decision support rather than continuous real-time control. Across the tested operating disturbances, COD removal after parameter adjustment remained between 88.7% and 94.2%. The contribution is the process-specific linkage between visual pollutant descriptors and constrained coagulation–flocculation parameter selection, not the isolated use of newer detection or segmentation backbones.