Scholarlink

<p>Predicting the concentrations of biological agents, such as bacterial cells and viral particles in water resource recovery facilities (WRRFs) is essential for monitoring emerging contaminants. Such predictive capabilities support efforts to protect public health and promote environmental sustainability. Recent data-driven models integrating a knowledge-based adaptation mechanism have been proposed to improve the generalizability of soft-sensor-based machine learning trained on unseen test datasets across wastewater matrices (WMs) and WRRFs. However, predicting biological contaminant concentrations in membrane bioreactors (MBRs) remains challenging due to the nonstationary nature of the process. Dynamic variations in water streams, data shifts across filtration layers, and complex interactions within biological communities introduce variability across datasets, potentially reducing prediction and generalization performances. We address this challenge by proposing a novel ensemble sequential data augmentation (EnSDA) model and lifelong learning-based knowledge adaptation to simultaneously predict bacterial cell and viral particle concentrations from available physicochemical input parameters across aerobic MBR (AeMBR) and anaerobic MBR (AnMBR) systems at two sites in Saudi Arabia. Lifelong learning with EnSDA improved and generalized bacterial and viral concentrations prediction accuracy and robustness on unseen MBRs while maintaining remarkable prediction performance across unseen WMs. We validated it by predicting the concentrations of bacterial cells, total virus, adenovirus, pepper mild mottle virus, coliphage, and CrAssphage across AeMBR and AnMBR plants using three different treatment facilities, and comparing the predictions with the Wasserstein generative adversarial network. Our work demonstrates that lifelong learning, combined with an efficient data augmentation method, simultaneously enables accurate and generalizable prediction of bacterial and viral pathogens from accessible WRRF measurements.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generalizable prediction of biological contaminants across aerobic and anaerobic membrane bioreactors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月2日

查看全文
  • Ibrahima N’Doye,
  • Jianxu Chen,
  • Yevhen Myshkevych,
  • Hasan Bagci,
  • Julie Sanchez-Medina,
  • Syed Haleem Shah,
  • Pei-Ying Hong

摘要

Predicting the concentrations of biological agents, such as bacterial cells and viral particles in water resource recovery facilities (WRRFs) is essential for monitoring emerging contaminants. Such predictive capabilities support efforts to protect public health and promote environmental sustainability. Recent data-driven models integrating a knowledge-based adaptation mechanism have been proposed to improve the generalizability of soft-sensor-based machine learning trained on unseen test datasets across wastewater matrices (WMs) and WRRFs. However, predicting biological contaminant concentrations in membrane bioreactors (MBRs) remains challenging due to the nonstationary nature of the process. Dynamic variations in water streams, data shifts across filtration layers, and complex interactions within biological communities introduce variability across datasets, potentially reducing prediction and generalization performances. We address this challenge by proposing a novel ensemble sequential data augmentation (EnSDA) model and lifelong learning-based knowledge adaptation to simultaneously predict bacterial cell and viral particle concentrations from available physicochemical input parameters across aerobic MBR (AeMBR) and anaerobic MBR (AnMBR) systems at two sites in Saudi Arabia. Lifelong learning with EnSDA improved and generalized bacterial and viral concentrations prediction accuracy and robustness on unseen MBRs while maintaining remarkable prediction performance across unseen WMs. We validated it by predicting the concentrations of bacterial cells, total virus, adenovirus, pepper mild mottle virus, coliphage, and CrAssphage across AeMBR and AnMBR plants using three different treatment facilities, and comparing the predictions with the Wasserstein generative adversarial network. Our work demonstrates that lifelong learning, combined with an efficient data augmentation method, simultaneously enables accurate and generalizable prediction of bacterial and viral pathogens from accessible WRRF measurements.