Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment
Dissolved oxygen (DO) control governs nitrification efficiency and aeration energy in activated sludge wastewater treatment. Physical DO sensors are subject to fouling, calibration drift, and the operational cost of multi-reactor instrumentation. A separate, structural constraint complicates soft-se...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
AUC Knowledge Fountain
2027
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1869483686467993600 |
|---|---|
| access_status_str | Open Access |
| author | Sorour, Abdelrahman Samir |
| author_browse | Sorour, Abdelrahman Samir |
| author_facet | Sorour, Abdelrahman Samir |
| author_sort | Sorour, Abdelrahman Samir |
| collection | Thesis |
| description | Dissolved oxygen (DO) control governs nitrification efficiency and aeration energy in activated sludge wastewater treatment. Physical DO sensors are subject to fouling, calibration drift, and the operational cost of multi-reactor instrumentation. A separate, structural constraint complicates soft-sensor design: the hardware-derived oxygen transfer coefficient exported by mechanistic simulators (here, K_La ≈ 120 d⁻¹ for the SUMO Tutorial Plant configuration used throughout) cannot independently close the operational DO mass balance once decoupled from the simulator's full kinetic model, because the parameter is calibrated inside a coupled biological-aeration formulation and is not portable as an isolated scalar.
This thesis develops a physics-informed neural network (PINN) soft sensor that frames the closure problem as an inverse identification of a lumped, plant-specific Effective K_La for a three-reactor aerobic cascade operating under standard SCADA inputs. The model accepts a 10-dimensional input vector — time, plus per-reactor temperature, influent flow, and aeration power — and produces simultaneous predictions of dissolved oxygen and a latent oxygen uptake rate (OUR). The architecture (approximately 142k trainable parameters) comprises a shared dense backbone with per-reactor DO and OUR heads, and is trained through a three-stage curriculum: (i) data-only baseline establishing the SCADA-to-DO mapping with frozen physics scalars, (ii) a Physics-Whisper Stage that introduces a weak physics loss (λ = 10⁻³) to drift the aeration scalars away from their initialization without disturbing the network, and (iii) a refinement stage that enforces strict mass-balance closure via differential learning rates between network weights and physics scalars.
The PINN identifies plant-specific Effective K_La values of 46.8, 49.5, and 50.7 d⁻¹ — lumped scalars that close the operational mass balance and fall within the literature range reported for fine-bubble aeration in process water. The unsupervised OUR latent matches SUMO's ASM-derived OUR within RMSE 35–65 mg L⁻¹ d⁻¹ for the buffered downstream reactors; the upstream reactor functions as a low-pass-filtered estimate of the high-frequency influent organic load (RMSE > 500 mg L⁻¹ d⁻¹). All OUR estimates are validated in silico against SUMO; physical respirometric validation is left for future work. A sensor topology analysis across four configurations shows that a bookend placement — DO sensors at the first and last reactors only — recovers the unmeasured interior reactor at RMSE 1.02 mg L⁻¹, within commercial DO probe tolerance, while forward simulation from a single upstream sensor exhibits unbounded drift (RMSE > 5 mg L⁻¹) consistent with an ill-posed initial-value problem.
Inference runs at approximately 11 ms per sample on consumer hardware — well below the 15-minute SCADA cycle and tractable for Nonlinear Model Predictive Control evaluation at SCADA cadence. The input-only architecture decouples inference from any specific DO probe used during training, although a deployed bookend topology still requires functional probes at the cascade boundaries. All training and validation data are generated by SUMO under a 365-day synthetic dynamic profile; sim-to-real validation against real-plant SCADA archives, together with respirometric validation of the OUR latent, are the immediate next steps before any deployment claim. |
| format | Thesis |
| id | oai:fount.aucegypt.edu:etds-3885 |
| institution | American University in Cairo (Egypt) |
| last_indexed | 2026-07-01T04:02:56.725Z |
| license_str | Not specified — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from AUC Knowledge Fountain — bepress |
| publishDate | 2027 |
| publishDateRange | 2027 |
| publishDateSort | 2027 |
| publisher | AUC Knowledge Fountain |
| publisherStr | AUC Knowledge Fountain |
| record_format | dspace |
| source_str | AUC Knowledge Fountain — bepress |
| spelling | oai:fount.aucegypt.edu:etds-3885 Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment Sorour, Abdelrahman Samir Dissolved oxygen (DO) control governs nitrification efficiency and aeration energy in activated sludge wastewater treatment. Physical DO sensors are subject to fouling, calibration drift, and the operational cost of multi-reactor instrumentation. A separate, structural constraint complicates soft-sensor design: the hardware-derived oxygen transfer coefficient exported by mechanistic simulators (here, K_La ≈ 120 d⁻¹ for the SUMO Tutorial Plant configuration used throughout) cannot independently close the operational DO mass balance once decoupled from the simulator's full kinetic model, because the parameter is calibrated inside a coupled biological-aeration formulation and is not portable as an isolated scalar. This thesis develops a physics-informed neural network (PINN) soft sensor that frames the closure problem as an inverse identification of a lumped, plant-specific Effective K_La for a three-reactor aerobic cascade operating under standard SCADA inputs. The model accepts a 10-dimensional input vector — time, plus per-reactor temperature, influent flow, and aeration power — and produces simultaneous predictions of dissolved oxygen and a latent oxygen uptake rate (OUR). The architecture (approximately 142k trainable parameters) comprises a shared dense backbone with per-reactor DO and OUR heads, and is trained through a three-stage curriculum: (i) data-only baseline establishing the SCADA-to-DO mapping with frozen physics scalars, (ii) a Physics-Whisper Stage that introduces a weak physics loss (λ = 10⁻³) to drift the aeration scalars away from their initialization without disturbing the network, and (iii) a refinement stage that enforces strict mass-balance closure via differential learning rates between network weights and physics scalars. The PINN identifies plant-specific Effective K_La values of 46.8, 49.5, and 50.7 d⁻¹ — lumped scalars that close the operational mass balance and fall within the literature range reported for fine-bubble aeration in process water. The unsupervised OUR latent matches SUMO's ASM-derived OUR within RMSE 35–65 mg L⁻¹ d⁻¹ for the buffered downstream reactors; the upstream reactor functions as a low-pass-filtered estimate of the high-frequency influent organic load (RMSE > 500 mg L⁻¹ d⁻¹). All OUR estimates are validated in silico against SUMO; physical respirometric validation is left for future work. A sensor topology analysis across four configurations shows that a bookend placement — DO sensors at the first and last reactors only — recovers the unmeasured interior reactor at RMSE 1.02 mg L⁻¹, within commercial DO probe tolerance, while forward simulation from a single upstream sensor exhibits unbounded drift (RMSE > 5 mg L⁻¹) consistent with an ill-posed initial-value problem. Inference runs at approximately 11 ms per sample on consumer hardware — well below the 15-minute SCADA cycle and tractable for Nonlinear Model Predictive Control evaluation at SCADA cadence. The input-only architecture decouples inference from any specific DO probe used during training, although a deployed bookend topology still requires functional probes at the cascade boundaries. All training and validation data are generated by SUMO under a 365-day synthetic dynamic profile; sim-to-real validation against real-plant SCADA archives, together with respirometric validation of the OUR latent, are the immediate next steps before any deployment claim. 2027-01-31T08:00:00Z thesis application/pdf https://fount.aucegypt.edu/etds/2820 https://fount.aucegypt.edu/context/etds/article/3885/viewcontent/abdelrahman_samir_sorour_thesis.pdf Theses and Dissertations AUC Knowledge Fountain Physics-Informed Neural Networks (PINNs) -Scientific Machine Learning (SciML)-Biological Wastewater Treatment-Virtual Soft Sensors-Dissolved Oxygen (DO) Control -Latent State Estimation-Inverse Parameter Discovery-Spatial Observability-Real-Time State Estimation Computational Engineering Construction Engineering and Management Environmental Engineering |
| spellingShingle | Physics-Informed Neural Networks (PINNs) -Scientific Machine Learning (SciML)-Biological Wastewater Treatment-Virtual Soft Sensors-Dissolved Oxygen (DO) Control -Latent State Estimation-Inverse Parameter Discovery-Spatial Observability-Real-Time State Estimation Computational Engineering Construction Engineering and Management Environmental Engineering Sorour, Abdelrahman Samir Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title | Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title_full | Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title_fullStr | Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title_full_unstemmed | Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title_short | Bridging Data and Dynamics: A Physics-Informed Soft Sensor for Dissolved Oxygen Control in BiologicalWastewater Treatment |
| title_sort | bridging data and dynamics a physics informed soft sensor for dissolved oxygen control in biologicalwastewater treatment |
| topic | Physics-Informed Neural Networks (PINNs) -Scientific Machine Learning (SciML)-Biological Wastewater Treatment-Virtual Soft Sensors-Dissolved Oxygen (DO) Control -Latent State Estimation-Inverse Parameter Discovery-Spatial Observability-Real-Time State Estimation Computational Engineering Construction Engineering and Management Environmental Engineering |
| url | https://fount.aucegypt.edu/etds/2820 https://fount.aucegypt.edu/context/etds/article/3885/viewcontent/abdelrahman_samir_sorour_thesis.pdf |
| work_keys_str_mv | AT sorourabdelrahmansamir bridgingdataanddynamicsaphysicsinformedsoftsensorfordissolvedoxygencontrolinbiologicalwastewatertreatment |