Leak Localization in Water Distribution Networks using Deep Learning

TLDR: This paper explores the use of deep learning for leak localization in Water Distribution Networks (WDNs) using pressure measurements by using a training data set including enough samples of all possible leak localizations, a Convolutional Neural Network can be used to learn the different pressure maps that carachterized each leak localization.

Abstract: This paper explores the use of deep learning for leak localization in Water Distribution Networks (WDNs) using pressure measurements. By using a training data set including enough samples of all possible leak localizations, a Convolutional Neural Network(CNN) can be used to learn the different pressure maps that carachterized each leak localization. The generalization accuracy has validated and evaluated by means of a testing data set. All of considered training, validation, and also testing data include leak size uncertainty, nodal water demand uncertainty and sensor noise. An innovative approach is proposed to convert every pressure residuals map to an image in order to apply a CNN. In addition with the purpose of filtering the effects of uncertainty and noise a time horizon Bayesian reasoning approach is used over each time instant classification output by the CNN. The Hanoi District Metered Area (DMA) is considered as a case study to illustrate the performance of the proposed leak localization method.

Figures

Fig. 2: Data generation scheme.

Fig. 8: Confusion matrix for sensor configuration 7 (12 sensors).

Fig. 1: Leak localization scheme using Deep Learning.

Fig. 3: Deep Learning Convolutional Neural Network architecture.

Fig. 4: Hanoi WDN.

Fig. 5: Training and testing flow patterns.

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Leak localization in water distribution networks using deep learning" ?

This paper explores the use of deep learning for leak localization in Water Distribution Networks ( WDNs ) using pressure measurements. The Hanoi District Metered Area ( DMA ) is considered as a case study to illustrate the performance of the proposed leak localization method. 

Q2. How many days have been generated for training and testing?

In particular, 30 daily flow patterns with a sample time of 10 minutes have been generated: 20 days for training data and 10 for testing data. 

Q3. What is the probable node in the network where the leak occurred?

Once np residuals r1, . . . , rnp associated to the pressure sensors installed in np inner nodes, an estimation of the all possible nn residuals associated to all the nodes of the network r̂1, . . . , r̂nn is computed using an interpolation method. 

Q4. What is the definition of the softmax layer?

The softmax layer is an activation function to normalize the FC layer output sum to one to utilize as a classification probability. 

Q5. What is the main problem of the proposed method?

The proposed method based on DL CNN should be able to localize any leak at any node despite of all possible sources of errors (such as demand estimation uncertainty, sensor noises and interpolation errors). 

Q6. How many different sensor configurations have been taken into account?

Seven different sensor configurations that consider a range from 4 to 12 pressure sensors installed in inner nodes of the network have been taken into account. 

Q7. What is the average value of residuals obtained in the data generation process?

As leak localization will be computed every hour and pressure values are computed every ten minutes, the average value of residuals obtained in data generation method presented in Fig. 2 are computed and provided to the Kriging and picture generation process. 

Q8. What is the purpose of the paper?

This paper proposes a leak localization method in WDNs that tries to exploit Deep Learning potentials for analysis and exploration through the map representation of pressure residuals of the WDN. 

Q9. What is the average leak noise of the EPANET system?

The profile of the 30 daily flow patterns is shown in Fig. 5.According to data generation method presented in Fig. 2, additive noise of ±5% has been added to pressure values computed by the hydraulic simulator that emulates the real WDN in the leak localization scheme depicted in Fig. 

Q10. What is the simplest way to compute the range of values of the raw pictures?

This action followed by a multiplication bound to set the range of values between 0 and 255, outcomes achieving a standard format of picture desired for the image classification processing. 

Q11. What is the time horizon for the leak localization task?

In order to improve the performance in the leak localization task, the leak localization task has been computed by means of (8) with different time horizons H = 1, ..., 24. 

Q12. What is the proposed method for detecting leaks in a WDN?

The proposed method follows the general scheme for online model-based Fault Detection and Isolation (FDI) strategies, where residuals are generated as the difference of actual sensor measurements (in this case pressure measurements in np inner nodes of the network S1, ..., Snp ) and the estimation of these values provided by a model of the WDN that considers no-leak conditions. 

Q13. What is the simplified model of the WDN of Hanoi?

This simplified model consists of one reservoir that supplies the inlet flow, 34 pipes and 31 inner nodes and its network graph is depicted in Fig.