knowledge transfer


Neuro-fuzzy approach to diagnosis of leakages and other operational faults in water distribution networks.

Submitted by Bogdan Gabrys 18.07.2002, IM


A lot of time, effort and resources is dedicated to purifying water but surprisingly high percentage of treated water is wasted through leakages in the water distribution networks (WDN) before even reaching customers. The occurrence of other operational faults like blocked pipes or erroneous states of valves etc. can also cause serious disruption in the services which need to be avoided. It is therefore very important that the state of the distribution system is continuously monitored. Unfortunately, due to financial constraints, it is not practical to measure all variables of interest and limited number of measurements are used together with WDN topology information to calculate the state of the WDN through state estimation procedures. Due to the scale of distribution systems an interpretation of the state can be quite a difficult task even for experienced system operators. Additionally if the so called topological error (e.g. leakage) occurs, the state estimation procedure usually results in a set of errors scattered across the network, making the diagnosis of the cause of the errors even more difficult. Though sequential analysis of precise numerical results of state estimation is useful, it also tends to ignore the greater picture of the overall system state which is something that experienced human operators primarily focus their attention on before analysing the detail. Therefore a neuro-fuzzy approach, thought to be mimicking the information processing and abstraction forming by human operators, has been proposed as a solution to this problem.


A neuro-fuzzy pattern recognition approach to fault detection and identification based on the examination of patterns of state estimates has been proposed to solve the problem. A General Fuzzy Min-Max (GFMM) neural network for clustering and classification has been used as a main building block in the developed recognition system. This hyperbox fuzzy sets based method has been designed to be able to process inputs in a form of confidence (real value) intervals, learn on-line, grow to meet the demands of the problem and include new information without need for retraining of the whole system, and cope with labelled and unlabelled data reflecting the fact that some of the network states are known (i.e. normal operating state etc.) while others are not. To improve the efficiency of the learning process and adaptability, the diagnostic system has been designed as a two-level hierarchical system. The first level consists of a GFMM neural network which selects one of the n second level “experts” (implemented using GFMM NN as well). In terms of water distribution systems the purpose of the first level of this recognition system is to distinguish different typical behaviour of the water system (i.e. night load, peak load etc.) while the second level components are responsible for the detection of anomalies for some characteristic load patterns. The second level can be, therefore, viewed as a collection of “load-pattern-experts”.

Status and results

The system has been implemented in Matlab and extensive simulation study have been conducted in which the two-level neuro-fuzzy recognition system has been trained and tested using data covering a 24-h period of operation of a realistic water distribution network. A set of 9144 training set representing 39 categories have been initially used. The 39 categories represented a normal operating state and leakages in 38 pipes of the network.The results have been tested on a separate testing set consisting of 91440 representing a mixture of normal operation and leakages (ranging from 0.002 to 0.029m^3/s) in all pipes of the network. The ability of GFMM to produce a graded response has been found invaluable in the process of restricting the area where the fault occurred if not enough accurate measurements are available to pinpoint the location of the leakage to a single pipe.The system has been also successfully tested for its ability to include information about new types of faults while in operation. The system which has been trained to recognise one type of anomaly (i.e. leakages) has been able to include new classes of operational faults (i.e. representing wrong status of valves and pipe blockages) utilising the GFMM ability to grow and adapt while in operation.

Adaptivity and portability

The adaptivity is an essential feature of the developed system. It can expand to include new information, has ability to adapt the parameters (hyperbox fuzzy sets) on-line and the hierarchical representation makes the approach even more flexible. Within the proposed two-level framework, distinctive variations in typical water network behaviour for different days of the week or seasons of the year can be quite easily accommodated by adding (removing) of the second layer experts and expanding (shrinking) the first layer network accordingly. Since the GFMM has been developed as a general purpose pattern classification/clustering approach it can be (and has been) applied to a number of different problems from different domains without any need for modifications making it also quite portable.

More information

Gabrys B. and Bargiela A., Neural Networks Based Decision Support in Presence of Uncertainties, ASCE J. of Water Resources Planning and Management, Vol. 125, No. 5, pp. 272-280, 1999.

Gabrys B. and Bargiela A., General Fuzzy Min-Max Neural Network for Clustering and Classification, IEEE Transactions on Neural Networks, Vol. 11, No. 3, pp. 769-783, 2000.

Gabrys, B., “Agglomerative Learning Algorithms for General Fuzzy Min-Max Neural Network”, a special issue of the Journal of VLSI Signal Processing Systems entitled "Advances in Neural Networks for Signal Processing", Vol. 32, No. 1/2, pp. 67-82., August-September 2002

Gabrys, B., “Neuro-Fuzzy Approach to Processing Inputs with Missing Values in Pattern Recognition Problems”, International Journal of Approximate Reasoning, in press, 2002

technical reports & case studies

executive summaries

scientific papers




search eunite with