News
For directions
see a map
of the campus and the route
to the university.
News | Team | Overview | Sensor development and deployment | Neural Network
A series of burst
trials is underway in collaboration with the University of East Anglia
and Yorkshire Water. The first set of trials took place in March 2001 and
the second set in May 2001. A further series is planned for July/ August
2001 and will include participation by the Water Engineering Research Unit
from Sheffield University.
The team presented
two papers at the 1st
IWA Conference on Instrumentation, Control and Automation, Malmo, Sweden,
June 3-7 2001. An html version of the powerpoint presentation for 'A
neural network approach to burst detection' is available here.
Team
Professor Torsun sadly died after a short illness on 12 November 2000. He will be remembered for his expertise and enthusiasm in temporal logic, intelligent knowledge based systems and neural networks, and for his academic and scientific contribution to the field. His work in the LAPS project has established a firm foundation for the successful conclusion of the Bradford research. Management of the Bradford project has been taken over by Professor Andrew Day and Dr. Alastair Wood, with the team of Steve Mounce, Asar Khan, and Peter Widdop. |
Prof. Andrew Day |
The principal objective of this research is to create an automated leakage
detection and location system based upon sensor technology and neural networks.
There are two main elements of the work. The first element is the development
of low cost failure sensors to supplement existing hydraulic instrument
data. The second element is the design of a Neural Network based system.
The neural network uses time series data produced by sensors to directly
construct an empirical model for predication and classification of leaks.
These components are being combined in a practical manner in an experimental
site in Yorkshire Water’s Keighley distribution system. Close collaboration
with Yorkshire Water has allowed an emphasis towards integration of these
developing technologies with existing system components (hydraulic sensors,
modelling software and existing data management procedures).
A data collection point in Keighley - photo taken during initial
inspection of Keighley sites
l to r: Asar Khan (UB), Steve Mounce (UB), John Machell (YW)
and Michael Campion (YW)
Yorkshire Water's Keighley distribution system provides water services
for around 26,000 customers. It contains approximately 110 km of pipeline
and seven service reservoirs. The distribution network has more recently
been partitioned into 15 DMAs. Six of these cover the city centre and immediate
environs and one of these (K710) was used as the experimental site for
the first phase of the project. A number of burst trials were carried out
in order to provide simulated data sets for the development of successful
neural network prototypes (during 1999 - 2000). More recently, circumstances
have brought about an unique opportunity to utilize an alternative DMA,
K709, as the test bed for a leakage detection case study. As well as newly
installed hydraulic sensors, the facilities were available to install the
new failure sensors developed and manufactured at Bradford University.
Hence during the year 2001, an extensive data set should be collected and
available for analysis.
Sensor development and deployment
The objective has been to research, design and develop, an appropriate failure sensor. It has to be cheap to manufacture, easy to install, and reliable in operation. It must afford sufficient accuracy, resolution and repeatability to be of use in the leakage detection and location system under development.
The Bradford research has produced two appropriate sensors intended to provide additional network performance information which might be indicative of leakage. They will provide a data input to the Neural Network system being researched and developed at Bradford. Two sensors have been developed:
1. The differential temperature failure sensor utilises the premise that the temperature of the soil surrounding a pipe with a leak or burst should only become close to the water temperature inside the pipe while water from inside the pipe continues to leak into the soil.
2. The opacity (cloudiness) failure sensor utilises the premise that local opacity or cloudiness of a water flow can be affected by the disturbance of sediment from elsewhere in the pipeline; such a disturbance may result from a sudden change in the flow regime, e.g. a sudden leak or burst.
The results obtained from the first prototype sensor are quite encouraging.
This has paved the way for the development of 10 similar prototype opacity
failure sensors. These have now been manufactured and installed for the
monitoring of abnormalities (burst, leakage and flushing) at selected
sites in DMA K709 in the Keighley area (as illustrated in the following
images taken winter 2000).
The opacity failure sensors back in the workshop.
The team apply appropriate safety precautions at the sample point
near Highfield Service Reservoir.
Access to the sample point requires the removal of heavy covers.
Once removed the access chamber is revealed.
The sensor housing and connecting blue valve.
John Machell ensures a sensor is sterilised prior to installation
The sensor is inserted into the pipe.
Once secure in the housing, the valve is turned back open.
Dr. Khan connects the wiring from a sensor to the data logger.
Peter Widdop installs a data logger into the roadside cabinet.
Neural Network and Data Analysis
The overall approach to leak detection and location is that of using measured data (time series produced by sensors) to directly construct an empirical model by use of one or more neural networks. The field of statistics concerned with analysing spatio-temporal data is called time series processing. In recent years, neural networks have been used for processing time series data. They have been shown to provide a powerful and elegant solution to problems that range from signal processing (speech recognition), prediction of stock market movements and the prediction of birth rate given that of previous years. The key concept in this case is that a neural network can learn to differentiate between a normal signature and an abnormal one, for some set of parameters which change in response to network events such as bursts. The theoretical underpinning of this (based on the state space model) has been researched and earlier work on static and time- delay neural networks successfully illustrated the validity of the approach for a simulated burst trial within a single DMA. However, for more general detection, it became apparent that due to both the inherent non- linear complexity of a water pipeline distribution network and also a number of practical limitations on the availability and quality of data, a robust approach has to be adopted capable of dealing with various granularities of scale. The neural network model fits this well with a distributed, data driven approach which is continually adjustable to it’s environment by retraining, and is good at handling noise and uncertainty. This system would need to be comprised of a number of modules which fit together in parallel as well as within a hierarchy. Three levels are needed:
Level 1: DMA level monitoring. Neural network detects to within a zone. Each zone level sensor modelled by one module outputting to a rule-based expert system which represents the architecture of a distribution system (written in collaboration with domain expert).
Level 2: Sub- zone monitoring within a DMA. This level relies on additional sensors been available within a DMA and possibly the use of hydraulic modelling software output. The experimental site described above will provide data sets for analysis at this level.
Level 3: Pipe level monitoring. The final level involves analysis of individual pipes, based on ‘virtual’ pipe level data from a hydraulic model. A very precise identification of leak location would rely on input to a neural network of signals from mobile acoustic sensor instrumentation.
Of course, the instrumentation present in a particular distribution system will dictate the depth in level possible.
A particular DMA level sensor (typically at the zone inlet) is modeled by one neural network which learns to forecast the sensor output. A number of architectures were investigated for forecasting, including static and recurrent networks (Elman and Jordan networks) in a conventional manner. However, review of the literature concerning Takens’ embedding theorem applied to forecasting resulted in an alternative approach. Of course, the main result of Takens is that we can predict future values of a time series solely on the basis of the observed past data without explicitly knowing the actual dynamics in the true system. However, when we have a stochastic time series the best we can do is predict the conditional probability which is the probability that we observe at the next time step t+1 given that we have observed the recent time series . Consequently neural networks have been applied to the task of the prediction of the entire conditional probability distribution of an unknown data generating process. A neural network model called the mixture density network (MDN) as presented by Bishop has been selected for development work. A prediction and classification detection module has been constructed using this model.
Current and future work involves extending the analysis to a more detailed
granularity of data (Levels 2 and 3 of the hierarchy), in particular the
output of the failure opacity sensors in order to refine the location process.
Main|Last
update: 31/01/01 | 
Steve
Mounce |