Water pollution is a major issue in the context of increasing population and industrialization, yet many drinking water treatment plants (DWTPs) are not fully efficient countering it. In particular, coagulation-settling stage often faces multiple disturbances and time lags, which lower the efficiency because coagulant dosage cannot be accurately calculated in real-time based on the effluent turbidity. To address this issue, we developed a method using deep learning image recognition to monitor the coagulation-settling stage in real-time. For that we used 5761 operational data and images of flocs from the sedimentation tank of a DWTP in East China in 2022, to build an image recognition regression model that predict the turbidity of the sedimentation tank effluent. Results show that our deep learning regression model, performs better with r-square (R²) of 0.97, mean absolute error (MAE) of 0.016 and mean absolute percentage error (MAPE) of 2.74%, compared with the traditional machine learning giving R² of 0.76, MAE of 0.045 and MAPE of 8.26%. The model also avoids misclassification at different turbidity intervals. The incorporation operational data of the sedimentation tank, prediction accuracy is improved by 79.6%. By adjusting the turbidity data to correct time misalignment, our model effectively handles the time lag caused by the hydraulic retention time of the sedimentation tank, thus enhancing the timeliness and accuracy of its practical application.

中文翻译：

---

图像识别增强了对饮用水处理厂混凝沉降的高效监测


在人口增长和工业化的背景下，水污染是一个主要问题，但许多饮用水处理厂 （DWTP） 并没有完全有效地应对它。特别是，混凝沉降阶段经常面临多重干扰和时间滞后，由于无法根据出水浊度实时准确计算混凝剂剂量，从而降低了效率。为了解决这个问题，我们开发了一种使用深度学习图像识别来实时监测凝血沉降阶段的方法。为此，我们使用了 2022 年华东地区 DWTP 沉淀池的 5761 个运行数据和絮凝体图像，构建了一个预测沉淀池污水浊度的图像识别回归模型。结果表明，与传统机器学习的 R 2 为 0.76、MAE 为 0.045 和 MAPE 相比，我们的深度学习回归模型在 R² 为 0.97、平均绝对误差 （MAE） 为 0.016 和平均绝对百分比误差 （MAPE） 为 2.74% 的情况下表现更好。该模型还避免了不同浊度区间的错误分类。沉淀池的结合运行数据，预测精度提高了 79.6%。通过调整浊度数据以纠正时间错位，我们的模型有效地处理了沉淀池水力停留时间引起的时间滞后，从而提高了其实际应用的及时性和准确性。