GB2367362A - Detecting leaks in water distribution systems - Google Patents

Abstract

A method of acoustic leak detection includes the construction of a noise power or amplitude signature. This may include the separation of the leak noise into a plurality of frequency range bands, and analysis to determine the scale of leak.

Description

DETECTING LEAKS IN WATER DISTRIBUTION SYSTEMS Field of the Invention This invention relates to the detection of leaks in water distribution systems.

Leaks in water distribution pipework generally emit acoustic noise. This phenomenon can be exploited in equipment for detecting leaks in buried pipes. Examples of such equipment range from simple listening sticks to sophisticated acoustic loggers and leak noise correlators. One method of detecting leaks acoustically is described in Patent Application No. WO 98/50771.

Although it is sometimes possible to locate leaks accurately, it is difficult to discriminate between large leaks and small leaks. Indeed, the loudest noises tend to be generated by the smallest holes, which may not warrant the expense of excavation and repair.

The problem of leak detection is further compounded by the following factors, as mentioned in our co-pending Application No.

0021358.7, to which reference should be made :- a) the pipework construction and materials affect the way in which acoustic noise is propagated-for example, noise in plastic pipes will sound quieter than the same noise in steel pipes, b) ambient noise, such as traffic noise, can obscure leak noise, and c) internal operation noise, such as that generated by customers drawing off leak noise, can obscure leak noise.

Co-pending Application No. 0021358.7 is thus concerned with the provision of an improved method and means for the detection of leaks which are such that an assessment of the likely size of a leak can be obtained.

The method described in said co-pending application includes dividing a multi-dimensional symptom space into a plurality of regions, carrying out a plurality of measurements which define a point in the symptom space and determining the relationship of this point to the diagnostic regions.

It is an object of the present invention to provide improvements in or modifications of the invention described in said co-pending Application No. 0021358.7.

Summary of the Invention According to a first aspect of the present invention there is provided a method of leak detection which includes the construction of a noise power or amplitude signature.

According to a second aspect of the present invention there is provided a means for detecting leaks which includes means for the construction of a noise power or amplitude signature.

A preferred method of construction and interpretation of the noise power or amplitude signature is described herein and includes the separation of the leak noise into a plurality of frequency range bands.

Brief Description of the Drawings Figure 1 shows the theoretical results of carrying out a leak noise correlation, Figure 2 shows a probable real life result of carrying out a leak noise correlation, Figure 3 shows the effects of applying correction factors to the graph of Figure 2, Figure 4 shows the separation of the noise emanating from a leak into three frequency bands, Figure 5 shows the results of correlation for one of said frequency bands, Figure 6 shows the results of correlation for all three frequency bands, Figure 7 shows a simple example of a neural network with four input conditions, Figure 8 shows an example of an artificial neuron evaluating a decision based on three input conditions, an Figure 9 shows how a typical input may be pre-classified.

Description of the Preferred Embodiment Patent Application No. 0021358.7, to which reference should be made, relates to the detection of leaks in water distribution networks and one area which it covers concerns leak sizing using statistical interpretation of acoustic noise signals in conjunction with external factors, such as pipe materials, pipe diameters, depths, etc. It refers to the use of leak noise correlators, which are instruments used to locate leaks precisely over short pipe sections, e. g. sections having a length of from 50 to 100 metres.

Application No. 0021358.7 also refers to how discrimination between leak noise and operational noise may be accomplished on acoustic loggers. It mentions splitting broad, all-encompassing frequency bands into narrower ones and carrying out a variety of noise measurements in each of these bands.

With leak noise correlators, current practice includes the use of two synchronised broad band signals, each being detected either side of the leak and these being appropriately filtered and crosscorrelated. The aim is to locate the position of the leak precisely.

Suppose there is a leak at a point in a buried pipe. One accelerometer or hydrophone is attached to the pipe (or a hydrophone inserted in it) at the nearest practical point to the suspected leak position. A second accelerometer or hydrophone is similarly placed on or in the pipe, but on the opposite side of the leak. The sensors will record random acoustic noise travelling up and down the pipeline from the leak and other unrelated sources. The correlator records the acoustic noise picked up simultaneously by both sensors. It carries out a cross correlation to determine the difference in arrival time T of the leak noise pattern at each sensor (see Figure 1). Knowing T and estimating or measuring the velocity of acoustic signals in the pipe enables an estimate to be obtained of the distance from one leak sensor to the leak.

An example of a prior art method of carrying out the above is that contained in PCT Application No. WO 98/50771. Considerable effort has been put into making the location of the leak as accurate as possible and the output of a leak noise correlator will typically be a graph of the cross correlation function. Ideally, this should show an unambiguous peak at the time lag corresponding to the position of the leak. However, there are many confounding problems, typically noise attenuation, echoes, etc. and the graph is more likely to be as shown in Figure 2. Methods have accordingly already been proposed which aim to eliminate the uncertainty and to produce unambiguous correlation plots as shown by the superimposed green trace in Figure 3.

These proposals have, however, not been totally satisfactory and the present invention is thus concerned with the calculation of noise power over different frequency ranges in a leak noise correlator-as described in detail below.

First, we record and store the raw broad band signals from the two sensors. Let us call this the raw data pair. We crosscorrelate this data, applying appropriate filtering, to give us the best estimate of T, i. e. the time lag between the two waveforms which gives us a strong unambiguous peak in the cross-correlation function due to the leak noise. This pinpoints the source of the leak noise in terms of the time lag T. Next, we note that the acoustic noise emanating from the leak can be regarded as made up of a range of frequencies (see Figure 4), which we may, for illustration purposes, separate into three bands, i. e. high, medium and low frequencies. Other and different numbers of frequency ranges may be used.

Taking each frequency range in turn, we filter the original raw data pair to generate a pair of sensor waveforms containing only those frequencies in one of the selected ranges. We cross-correlate this data pair and note the amplitude of the correlation at the time lag T. This amplitude will be proportional to the power of the acoustic leak noise received by the two sensors at the selected frequency range (see Figure 5). It may or may not be related to the size of the leak itself.

The process is repeated for each frequency band, ending up with a set of power values covering all the frequency ranges, as shown in Figure 6. We can call this the leak noise"power signature", which can be analysed in conjunction with external factors, such as pipe size, material, pressure, etc. , to estimate the leak size statistically based on previous experience. Leak size classifiers which may be used for this purpose include Naive Bayes classifiers, Decision Tree classifiers and support vector/kernelbased classifiers.

When constructing the power signature, each individual power value may be obtained by averaging cross-correlation values in a localised region around the location time lag T. This may mitigate effects of noise on the original estimate of T.

Instead of a power signature, an amplitude signature may alternatively be constructed, each amplitude being the square root of the cross-correlation function at or averaged around T for each frequency range.

It is also possible to investigate relationships between individual components or groups of components of the power or amplitude signature in isolation or in combination with external factors such as pipe material, pressure or other factors and relate these to leak size. For example, the relative strength of one power relative to the combined power of all the other components may be an important indicator of leak size.

If a larger number of frequency bands are used and the crosscorrelation function computed for time lag T at each of these, it is possible to calculate statistical moments of the distribution of the power values and examine the relationship of these in isolation, or in combination with external factors such as pipe material, pressure or other factors, to leak size.

As an alternative to the use of statistical relationships, a neural net approach may be employed, i. e. a system is modelled by an interconnecting matrix of neurons. Each neuron receives one or more inputs from other neurons, or signals from the outside world called the input layer, in this instance, values of noise level, pipe type, etc. Each neuron assigns weights to each of its inputs and, in a simple system, sums the weighted inputs to produce an output value. This output value is presented to a"threshold unit"which fires a signal to the next neuron in the chain if the output value of the first neuron exceeds a predetermined limit. Ultimately, after a sequence of such neural signal transmissions, values are accumulated at nodes on an output layer.

Five output layer nodes may be provided, for example, tiny leak, small leak, medium leak, large leak and huge leak. The output node with the largest accumulated value is considered to be the most likely diagnosis for the particular set of inputs presented to the network. Figure 7 shows a simple example of such a network with four input conditions.

The neural network must be trained to activate the right output for a given set of input conditions. In particular, values must be calculated for the weighting factors and threshold functions used at the intermediate nodes of the network. A typical set of weights and a threshold function are illustrated in Figure 8 which shows an example of an artificial neuron evaluating a decision based on three inputs for pipe material characteristic, water pressure and noise feature.

To calculate the weighting factors and threshold functions, the network is presented with a training set of various input conditions with known correct outputs. Thus, the training sets could comprise different combinations of noise measurements, pipe type, pressure, etc. , with each set having a known diagnostic output, i. e. a known leakage size.

Various established algorithms are known for recursively adapting the weights and the threshold levels on the intermediate neuron layers so that correct outputs are generated for each of the given set of input conditions.

In a refinement of the neural net approach to leak size evaluation,"fuzzy"classification procedures are incorporated in the measurands presented at the inputs to the network and to the outputs calculated by the network. This can simplify processing of the information.

Thus, as an example, Figure 9 shows how a typical input for pressure may be pre-classified into one of three"fuzzy"categories of low, medium and high pressure before being presented to the neural network.

Claims (7)

Claims : -

1. A method of leak detection which includes the construction of a noise power or amplitude signature.

2. A method of leak detection as claimed in Claim 1, in which construction of the noise power or amplitude signature includes the separation of the leak noise into a plurality of frequency range bands.

3. A method of leak detection as claimed in either of the preceding claims, carried out substantially as hereinbefore described with reference to the accompanying drawings.

4. Means for detecting leaks which includes means for the construction of a noise power or amplitude signature.

5. Leak detecting means as claimed in Claim 4, which includes means for interpretation of the noise power or amplitude signature.

6. Leak detecting means as claimed in Claim 5, in which the means for construction and interpretation of the noise power or amplitude signature includes for effecting separation of the leak noise into a plurality of frequency range bands.

7. Leak detecting means as claimed in any one of Claims 4 to 6, constructed and arranged to operate substantially as hereinbefore described with reference to and as shown in the accompanying drawings.

Calculate difference in arrival time T of leak noise pattern at each sensor (eg B detects common noise pattern T secs after A) Assume noise propogates in both directions at constant velocity V Thus Extra distance travelled by noise to sensor B is VT Thus Distance to leak from A is L = 0. 5 (D-VT) Fui. 1