Systems for pipeline leakage detection

Leak detection carried out by water companies is focused on analysis of

night flow measurements in district metering areas. The measurements are

almost always manual and use a number of methods including leak correlation,

step testing and the use of electronic listening sticks. These methods,

apart from noise correlation, are time consuming, labour intensive and

unreliable.

– to investigate the causes of failure of buried infrastructure through the

collection and analysis of historical data;

– to develop a neural network and statistical computer system which will

continuously, automatically and centrally monitor a pipeline distribution

network;

– to refine the failure detection system to include a component for the

prediction of pipeline bursts and water quality and computing consequences

of failure and cost benefit analysis of pipe replacement; and

– to disseminate and exploit the research output effectively.

The project brings together mathematical and theoretical foundation,

advanced computational and statistical models, engineering, management and

economic expertise within the consortium.

Signal capture and low-cost data acquisition present major challenges, being

required for all the signals collected. The monitoring systems envisaged

need data from a number of sensors distributed over the network to provide

snapshots of its state. A proper snapshot requires data from each sensor to

be time stamped so that the network can be viewed as a whole at

giveninstants. High resolution time-stamping of data is therefore a high

priority. An initial study of this problem indicates that use of the Rugby

radio time signal may provide a solution.

As well as collecting data, the planned active network sounding methods

introduce an acoustic signal into the pipe that can be detected and acquired

at other locations within the locale. The low level acoustic signal

generated will be carefully crafted, based on pseudo-random binary sequences

and is designed to enable the acoustic propagation characteristics of the

pipe network to be determined. The hardware needed to generate and inject

these signals will be integrated with the data acquisition systems.

Data obtained using the acquisition systems will be stored in a raw binary

form so that future processing schemes can be applied to the original data

if necessary. A second part of the data collection activity will be the

design and development of a temporal-spatial database to store pipeline

signatures and associated feature sets.

Historical data (hydraulic and water quality) obtained from Yorkshire Water,

Anglian Water and Thames Water, as well as on-line dynamic data generated

from Yorkshire Water will provide sufficient volume of data to characterise

the overall behaviour of a pipeline distribution network (pipeline

signature). In addition, correlation-based leak detection using acoustic

methods will be used. Examples include the use of knowledge about the

acoustic transmission properties of the pipe to compensate for frequency

dispersion effects in the evaluation of cross-correlation functions. It is

known for studies at the University of East Anglia that uncompensated

dispersion can severely impair the detection of leaks by this method.

Another example is the use of adaptive filtering methods coupled with neural

net pattern processing to remove extraneous interfering noise from the

acoustic signals used during cross correlation which is already known to be

a problem in plastic pipes. Both methods will be developed during the

project.

The signature observed on a network will be used as an indicator of normal,

fault-free or abnormal operation. Discrimination between normal and abnormal

signatures will be analysed using neural nets and statistical methods. In

the case of neural nets signature analysis, unsupervised self-organising an

supervised neural networks will be cascaded to learn and extract the salient

features of the data in normal and abnormal conditions. In the case of

statistical signature analysis, the probability distribution of observed

signatures is estimated and used to assess the likelihood of a particular

observation.

Techniques to allow the neural network to Œlearn¹ by being exposed to

acceptable and unacceptable patterns of signals will be researched and

optimised. Included in this part of the research will be the effect of

natural variations (eg diural) and long-term pipeline degradation effects.

The features used in the signature will be actual values of normalised

pressure and flow coupled with a compact feature of the acoustic signal such

as a correlation function. However, these features must be fused together in

a principled manner such that discriminatory features are not overpowered by

less discriminatory or noisy features. Simple approaches (eg variance

normalisation, principal component analysis) are strengthened by the use of

class discriminatory transformations of the feature using both analytic and

neural net methods. In both neural net and statistical approaches to

signature analysis, rigorous methods or statistical cross validation will be

used to ensure that results are meaningful.

Interpretation of pipeline signature will be carried out using neural

networks and statistical computing models. Neural networks architecture will

be determined from the nature of the problem to be solved which includes

monitoring requirements (water leaks, pipe deformation, joint failure,

cracks, obstructions etc) and pipeline signature. Initially research will

focus on a temporal adaptive neural network. This type of neural network is

able to learn without being given correct answers for the input pattern. It

is effective with problems whole algorithms are too complex to define, and

is able to compensate for accuracy and noise in sensor readings. It offers a

method for dealing with unexpected and changing situations and allows for

incremental learning. These properties match favourable with the irregular,

noisy and unexpected pattern of sensor signals we envisage obtaining from

the pipeline. The neural network will be trained with Œseen’ data to ensure

its stability and convergence. Random faults will be generated in the

pipeline, and then the network will be tested with unseen data to determine

the distributive effect of one or more Œfaults’ propagating through the

pipeline. Statistical methods will be used to interpret and validate

acoustic signal and the neural network.

While a neural network is a powerful recognition and learning technique, it

can neither generate actions from the knowledge it has learned nor can it

interpret the knowledge. An effective method to solve this problem is to

generate rules from knowledge learned by the neural network and then use

these rules to construct a knowledge-based expert system that can interpret

data, generate actions and give explanations. The University of Bradford

team is developing a system which automatically generates a rule-set from a

neural network. It uses an optimisation algorithm based on cluster analysis

to eliminate redundant rules. It is proposed to further develop the system

to generate a rule-set from the neural network, which can be used to provide

recommendations, actions and explanations of pipeline behaviour.

The approach of network signature analysis is complementary to the methods

proposed by the consortium teams at Brunel University and Imperial College.

These are essentially based on Œstrong’ analytic models of the network,

whereas those of University of East Anglia and University of Bradford are

based on Œreal¹ data models that are based on large amounts of training data

taken from the network.

The efficient management of leakage in water distribution networks is high

on the agenda of the UK water industry. This research proposal has been met

with enthusiastic response from the collaborating water companies because it

addresses problems which are currently either impossible or very difficult

to solve successfully.

The programme of research is designed to generate a series of incremental

deliverables that will allow the industry and consumers to benefit both in

the short and long term. For example, a short-term deliverable will be an

enhanced acoustic correlation system working well on plastic pipes, other

materials and trunk mains. The more ambitious concepts of local network

signature analysis is a medium-term goal that may allow leaks to be detected

remotely over a wide area and has the potential to save water and reduce

operating costs. The concept of a highly instrumented pipe installation

using low cost sensors and efficient data capture is a long-term goal that

would enable failures to be predicted before they occur. This would be of

great benefit in saving water and reducing costs to consumers and companies

alike.

In addition to these tangible goals, the research is likely to lead to the

development of a family of signal processing and signature analysis

techniques that may find application in domains other than the water

industry, such as the petrochemical, gas and power distribution networks.