Robust leak localization in water distribution networks using machine learning techniques
- Autores: Adrià Soldevila Coma
- Directores de la Tesis: Vicenç Puig (dir. tes.), Sebastián Tornil Sin (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2018
- Idioma: español
- Tribunal Calificador de la Tesis: Ramón Pérez Magramé (presid.), Tom Nørgaard Jensen (secret.), Daniela Fuchs Hanusch (voc.)
- Programa de doctorado: Programa de Doctorado en Automática, Robótica y Visión por la Universidad Politécnica de Catalunya
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
This PhD thesis presents a methodology to detect, estimate and localize water leaks (with the main focus in the localization problem) in water distribution networks using hydraulic models and machine learning techniques. The actual state of the art is introduced, the theoretical basis of the machine learning techniques applied are explained and the hydraulic model is also detailed. The whole methodology is presented and tested into different water distribution networks and district metered areas based on simulated and real case studies and compared with published methods.
The focus of the contributions is to bring more robust methods against the uncertainties that effects the problem of leak detection, by dealing with them using the self-similarity to create features monitored by the change detection technique intersection-of-confidence-interval, and the leak localization where the problem is tackled using machine learning techniques. By using those techniques, it is expected to learn the leak behavior considering their uncertainty to be used in the diagnosis stage after the training phase.
One method for the leak detection problem is presented that is able to estimate the leak size and the time that the leak has been produced. This method captures the normal, leak-free, behavior and contrast it with the new measurements in order to evaluate the state of the network. If the behavior is not normal check if it is due to a leak. To have a more robust leak detection method, a specific validation is designed to operate specifically with leaks and in the temporal region where the leak is most apparent.
A methodology to extent the current model-based approach to localize water leaks by means of classifiers is proposed where the non-parametric k-nearest neighbors classifier and the parametric multi-class Bayesian classifier are proposed.
A new data-driven approach to localize leaks using a multivariate regression technique without the use of hydraulic models is also introduced. This method presents a clear benefit over the model-based technique by removing the need of the hydraulic model despite of the topological information is still required. Also, the information of the expected leaks is not required since information of the expected hydraulic behavior with leak is exploited to find the place where the leak is more suitable. This method has a good performance in practice, but is very sensitive to the number of sensor in the network and their sensor placement.
The proposed sensor placement techniques reduce the computational load required to take into account the amount of data needed to model the uncertainty compared with other optimization approaches while are designed to work with the leak localization problem. More precisely, the proposed hybrid feature selection technique for sensor placement is able to work with any method that can be evaluated with confusion matrix and still being specialized for the leak localization task. This last method is good for a few sensors, but lacks of precision when the number of sensors to place is large. To overcome this problem an incremental sensor placement is proposed which is better for a larger number of sensors to place but worse when the number is small.
