Morphology-informed deep learning for risk assessment of filamentous bulking in a full-scale industrial wastewater treatment plant.
BORZOOEI, Sina; SCABINI, Leonardo Felipe dos Santos; ZHU, Jun-Jie; DANESHGAR, Saba; DEBLIECK, Lukas; BROECK, Elias Van Den; TORFS, Elena.
BORZOOEI, Sina; SCABINI, Leonardo Felipe dos Santos; ZHU, Jun-Jie; DANESHGAR, Saba; DEBLIECK, Lukas; BROECK, Elias Van Den; TORFS, Elena.
Abstract: Filamentous bulking (FB) is a recurring cause of solids-separation failure in activated sludge systems. Standard diagnostics such as sludge volume index (SVI) and routine microscopy are largely reactive, confirming deterioration only after settling performance has already degraded. Surrogate, rule-based risk assessment models offer interpretable early warnings, but they infer microbial states indirectly and may become unreliable under variable industrial operation. Here, we propose a deep-learning, morphology-informed framework that directly scores FB risk from sludge microscopy images. Two pretrained computer vision (CV) models (ConvNeXt-nano and CrossViT-18) were fine-tuned using a two-year image dataset from a full-scale industrial hybrid feast-famine sequencing batch reactor (SBR) wastewater treatment plant (WWTP). Although trained with binary image labels, both models produced continuous probability outputs that reflected gradual morphological transitions. Daily risk scores were obtained by aggregating image-level predictions and evaluated against operator-confirmed FB episodes. Performance was benchmarked against a site-adapted fuzzy logic risk assessment model (RAM) driven by surrogate process indicators. The RAM showed limited episode-level sensitivity (<30% across episodes) and a higher false positive rate during non-bulking periods (21.7%). In contrast, both CV-based models yielded more consistent episode tracking. CrossViT-18 achieved episode-level sensitivities of up to 92% while exhibiting no false-positive alerts during confirmed non-bulking periods, whereas ConvNeXt-nano achieved 97.6% non-bulking accuracy with a 2.4% false positive rate. These results support microscopy-driven risk scoring as a practical complement to surrogate-based early-warning logic for proactive sludge monitoring in industrial WWTPs.
@article={003300501,author = {BORZOOEI, Sina; SCABINI, Leonardo Felipe dos Santos; ZHU, Jun-Jie; DANESHGAR, Saba; DEBLIECK, Lukas; BROECK, Elias Van Den; TORFS, Elena.},title={Morphology-informed deep learning for risk assessment of filamentous bulking in a full-scale industrial wastewater treatment plant},journal={Environmental Research},note={v. 301, p. 124520-1-124520-12 + supplementary data},year={2026}}