DETECTION AND LOCALIZATION OF WATER LEAKS IN WATER NETS BY MEANS OF A MONITORING SYSTEM, HYDRAULIC MODEL AND NEURONAL NETWORKS

Transcription

1 ESM 2013 Lancaster DETECTION AND LOCALIZATION OF WATER LEAKS IN WATER NETS BY MEANS OF A MONITORING SYSTEM, HYDRAULIC MODEL AND NEURONAL NETWORKS Jan Studzinski, Izabela Rojek Polish Academy of Science

2 Agenda INTRODUCTION CALCULATION APPROACH CALCULATION RESULTS CONCLUSIONS

3 INTRODUCTION 1/4 At the Systems Research Institute of Polish Academy of Sciences (IBS PAN) an ICT system for computer aided management of communal water networks is for a couple of years under development One of its tasks is the water net detection and failures localization The water losses that result from the water net failures can amount to 30% of the total water production in the waterworks They cause considerably financial losses in the enterprise The reduction of these losses is an essential help in the management of waterworks 3

4 INTRODUCTION 2/4 In the algorithm presented the hydraulic model of the water net is treated as the real object and using it some water net models in form of the MLP and Kohonen neuronal nets are developed The algorithm has been tested on the real data obtained from the Polish Waterworks in Rzeszow and it will be included into the ICT system for the water networks management developed at IBS PAN 4

5 INTRODUCTION 3/4 The algorithm consists of the following steps: 1. Defining the ranking list of the sensitive points in the water net 2. Choice of the most sensitive points for the SCADA system planned 3. Calculating the distribution of the standard pressure and flow values defined on the chosen measurement points by means of the water net hydraulic model 5

6 INTRODUCTION 4/4 4. Calculating the pressure and flow distributions for the simulated failure states using the water net hydraulic model 5. Developing the classifier for failures localization in form of MLP and Kohonen 6. On-line registration of the current pressure and flow values from the SCADA measurement points 7. In case of an inadmissible measurement finding out from the developed neuronal nets such the model that indicates the water net node with the presumable water leak

7 CALCULATION APPROACH 1/9 The first step of the algorithm means defining the most sensitive points of the water net that could be chosen as the measurement points for the planned SCADA system These points can be identified with the formulas in which the impact of simulated water leaks on the changes of water pressure and flow values in different water net nodes is calculated The nodes with the biggest value changes are chosen for the SCADA measurement points A program for planning the SCADA system has been developed at IBS PAN and included into the referred ICT system All the calculations testing the algorithm have been done with hydraulic model of the water net 7

8 CALCULATION APPROACH 2/9 The investigated water network 8

9 CALCULATION APPROACH 3/9 The second step of the algorithm is the choice of the measurement points that will be used for defining the standard and failure distributions of water pressures and flows The algorithm presented tested only the water net nodes sensitive against the flow changes The investigation have been considered for 10 and 20 measurement points The water leaks were simulated in 37 nodes in the first case of investigation and in 44 nodes in the second one There are 390 nodes in the water net investigated 9

10 CALCULATION APPROACH 4/9 The third step of the algorithm is the calculation of the hydraulic model to determine the flows distribution in the measurement points for the standard water load of the water net An appropriate hydraulic model of the water net has been developed at IBS PAN and included into the referred ICT system 10

11 CALCULATION APPROACH 5/9 The fourth step of the algorithm is the calculation of the simulated failure cases in the water net using its hydraulic model In the given nodes of the water net the water leaks are simulated and the resulting distributions of the flow values in the measurement points are recorded In each distribution the measurement point with the biggest reaction on the water leak is identified In this step the learning files for the MLP neuronal nets used in the following steps are prepared. 11

12 CALCULATION APPROACH 6/9 Part of data file for teaching the neuronal nets 12

13 CALCULATION APPROACH 7/9 The fifth step of the algorithm means the development of the classifier to signalize the failure state and to indicate its localization in the water net The classifier is in form of a neuronal net which is created in the classical way on the basis of three data files for teaching, testing and validation The main problems while realizing this step are the preparation of the learning files and the choice of the neuronal model structure In the latter case the number of layers for the neuronal net, the number of neurons situated on the layers and the transition functions between the layers have to be established in advance 13

14 CALCULATION APPROACH 8/9 The error functions (stop criteria) used at the teaching process of the neuronal nets are optionally the sum of squares function (SOS) and the cross entropy The formula for SOS is: where N is the number of examples used, yi is the consecutive calculated output value of the neuronal net and ti is the real value from the data file The cross entropy function is:

15 CALCULATION APPROACH 9/9 The sixth and seventh steps of the algorithm are conducted after the classifier for identification of the failure states on the water net is already made The reliable execution of this task depends strictly on the quality of the SCADA system installed on the water net and of the classifier constructed 15

16 CALCULATION RESULTS 1/9 The first classifier models have been created in form of the MLP neuronal nets with one hidden layer By the learning experiments 2 parameters have been optionally changed: - the number of neurons on the hidden layer from 5 up to 30, and - the number of the learning runs that took the values 200, 500 and

17 CALCULATION RESULTS 2/9 The inputs of the neuronal net investigated are the water flow values in 37 or 44 nodes of the water net obtaining from the hydraulic model The net output is the number of the SCADA measurement point that was the most sensitive against the simulated water leak By all learning runs the teaching file consisted of 70% of all examples used and the testing and validation files amounted to 15% of the examples In the calculation the number of all examples used was Technical Exchange

18 CALCULATION RESULTS 3/9 The MLP nets calculated for 20 monitoring points 18

19 CALCULATION RESULTS 4/9 The qualities of the MLP nets obtained for 20 monitoring points in % 19

20 CALCULATION RESULTS 5/9 In case of the Kohonen nets their parameterization was made by changing the number of teaching runs from 1000 to Kohonen nets calculated for 20 monitoring points 20

21 CALCULATION RESULTS 6/9 The errors of the Kohonen nets obtained for 20 monitoring points in % 21

22 CALCULATION RESULTS 7/9 The best MLP and Kohonen nets calculated for 20 monitoring points 22

23 CALCULATION RESULTS 8/9 The network qualities for the best MLP and Kohonen nets calculated for 20 monitoring points in % 23

24 CALCULATION RESULTS 9/9 The best MLP and Kohonen nets calculated for 10 and for 20 monitoring points 24

25 CONCLUSIONS 1/6 By the subsequent teaching runs of a neuronal net the calculation error is gradually sinking and the error for the parallel conducted testing runs begins from a moment to grow It means that the teaching process shall be stopped to avoid the over-fitting of the neuronal model The event of over-fitting depends on the number of neurons on the hidden layer of the net: - a small number of hidden neurons results in an inability of the neuronal net to extract the sufficient knowledge of the problem solving from the data files - many neurons on the hidden layer cause that the net learns the teaching data to exactly and as a result it can not then to generalize its knowledge with other data than the teaching ones 25

26 CONCLUSIONS 2/6 The calculation results show that in the experiments with 20 measurement points the MLP neuronal nets with the best ability to recognize and localize the water leaks in the water net are: MLP with the cross entropy function as the error function, with the Tanh transition function on the hidden layer and with the Softmax function on the output layer MLP with the cross entropy function as the error function, with the Logistic transition function on the hidden layer and with the Softmax function on the output layer MLP with the cross entropy function as the error function, with the Tanh transition function on the hidden layer and with the Softmax function on the output layer 26

27 CONCLUSIONS 3/6 The quality values of these MLP nets are 100%, 100% and 98%, respectively. In these classifiers as the error function the cross entropy function was used and as the activation functions on the hidden and output layers the combinations of functions Tanh-Softmax or Logistic- Softmax appeared to be best. Other function combinations like Tanh-Logistic or Logistic-Linear or Exponential-Exponential were unsuccessful. 27

28 CONCLUSIONS 4/6 As the best Kohonen classifier appeared the network SOFT with 100 neurons on the topological layer (10x10) and with the simulation runs number of 1000 that has been calculated for 20 monitoring points. Its quality value equals to 75,51%.

29 CONCLUSIONS 5/6 In general the MLP classifiers are better than the Kohonen ones although the Kohonen nets are more complicated than these of MLP. In case of MLP nets the number of monitoring points considered has got an essential impact on the detection and localization of the water leaks. The classifier MLP is essentially worse than MLP although the number of damaged water net nodes that it had to detect was smaller than in this another case. This observation does not concern the Kohonen nets that do not fit in general for solving the problems of finding out the water net failures.

30 CONCLUSIONS 6/6 With the algorithm presented the hidden water leaks in the water net can be fast and correct discovered and localized and in this way the water as well as the financial losses in the waterworks can be reduced Another way of reducing these losses is the computer aided generation of optimal plans for conducting the revitalization works on the water net In IBS PAN the appropriate programs for generating the water net revitalization plans are under development using the fuzzy sets algorithms With both kinds of programs: for localizing the water net failures and for planning the water net revitalization, the communal waterworks can receive the complex and innovative information tools for a better management of the enterprises 30

31 The presentation has been realized in frame of the research project of the Polish National Centre for Research and Development (NCBiR) co-financed by the European Union from the European Regional Development Fund, Sub-measure "Development Projects"; project title: IT system supporting the optimization and planning of production and distribution of water intended for human consumption in the subregion of the central and western province of Silesia ; project ref no POIG /12.

32 Thank you for your attention