To address the difficulty of directly sensing in-tank sedimentation states during sludge discharge in horizontal-flow sedimentation tanks (HSTs), this study proposes a soft-sensing framework for bottom-sludge thickness in drinking water treatment plants. This framework is designed to overcome the limited capacity of effluent-turbidity-based indicators for fine-grained discharge control and the impracticality of applying computational fluid dynamics (CFD) to real-time state estimation. The framework integrates Supervisory Control and Data Acquisition (SCADA) operational data, ultrasonic sludge–water interface measurements, and CFD-derived hydraulic priors. To incorporate hydrodynamic knowledge of sediment-particle transport, three fusion paradigms are developed: parameter transfer, representation fusion, and knowledge distillation, injecting physical priors into the parameter space, latent representation space, and supervision-constraint space, respectively. Performance is evaluated using pointwise accuracy (PA), curvature consistency error (CCE), and mass-conservation error (MCE). Experiments on a real-world HST dataset show that, across the six predictors examined, the three paradigms reduced PA, CCE, and MCE by 30.7%, 16.0%, and 56.3% on average relative to the same predictors trained without prior fusion. Under the in-distribution setting, the Attention predictor combined with parameter transfer attained the lowest PA (0.026) and the lowest MCE (1.052) among the eighteen paradigm–predictor combinations evaluated. Under the out-of-distribution setting with extended sedimentation duration, knowledge distillation attained the lowest values on all three metrics across zero-shot, 4-shot, and 6-shot adaptation; in the zero-shot setting, its PA, CCE, and MCE were 33.3%, 50.9%, and 33.8% lower than those of the second-best paradigm. These results demonstrate, within the experimental scope of this study, a methodological foundation for state-informed sludge-discharge scheduling in HSTs.
Meng et al. (Thu,) studied this question.