Keywords
Water distribution network assets analysis; GPR image interpretation; Intelligent data analysis;Multi-agent systems; Semi-automatic labelling of GPR images
Start Date
7-7-2022 2:00 PM
End Date
7-7-2022 2:20 PM
Abstract
Concerns around climate change, disaster resilience, and digitization underscore the importance of reliable and affordable information about critical infrastructure such as water distribution networks (WDNs) in order to prevent/mitigate potential negative effects. For instance, lack of data from WDNs enables problems such as leaks to persist undetected. While manual and destructive detection methods exist to monitor the health of WDNs (e.g., impact echo, laser scan, among others), these methods lack the power to provide comprehensive data for continuous modelling, operation, assessment, and maintenance. Ground penetrating radar (GPR) offers one non-destructive testing method that is inexpensive and easily operable; however, the challenge to GPR lies in image interpretation. At present, no consistent protocol for the interpretation of GPR data exists, and thus issues such as noise and widely varying testing conditions block the incorporation of GPR data into WDN management. Thus, reliable methods for extracting WDN information from raw GPR images are essential. This work extends methods of GPR data analysis to extract the features of interest (e.g., pipes and leaks) from raw GPR images to increase their interpretability. While methods comparing current/initial state images have been successful, initial state data is not generally available. Therefore, this research identifies methods of data manipulation that will reliably extract features from raw images without the need for comparison. Multi-agent techniques have been used in this paper to break the GPR data into groups of features, or families, including both pipes and leaks of interest and noise. The crux of this work is to identify the geometric attributes that can distinguish these families of interest from the noise. This process of identification of the distinguishing attributes of families of interest advance toward a reliable process for automatic data mining of raw GPR images to produce interpretable images for operators of WDNs.
Improving Water Distribution Network Data Extraction from RawGPR Images with Multi-Agent Techniques and Machine Learning
Concerns around climate change, disaster resilience, and digitization underscore the importance of reliable and affordable information about critical infrastructure such as water distribution networks (WDNs) in order to prevent/mitigate potential negative effects. For instance, lack of data from WDNs enables problems such as leaks to persist undetected. While manual and destructive detection methods exist to monitor the health of WDNs (e.g., impact echo, laser scan, among others), these methods lack the power to provide comprehensive data for continuous modelling, operation, assessment, and maintenance. Ground penetrating radar (GPR) offers one non-destructive testing method that is inexpensive and easily operable; however, the challenge to GPR lies in image interpretation. At present, no consistent protocol for the interpretation of GPR data exists, and thus issues such as noise and widely varying testing conditions block the incorporation of GPR data into WDN management. Thus, reliable methods for extracting WDN information from raw GPR images are essential. This work extends methods of GPR data analysis to extract the features of interest (e.g., pipes and leaks) from raw GPR images to increase their interpretability. While methods comparing current/initial state images have been successful, initial state data is not generally available. Therefore, this research identifies methods of data manipulation that will reliably extract features from raw images without the need for comparison. Multi-agent techniques have been used in this paper to break the GPR data into groups of features, or families, including both pipes and leaks of interest and noise. The crux of this work is to identify the geometric attributes that can distinguish these families of interest from the noise. This process of identification of the distinguishing attributes of families of interest advance toward a reliable process for automatic data mining of raw GPR images to produce interpretable images for operators of WDNs.
Stream and Session
false