A leakage detection system extracting the most meaningful features with decision trees.
- 1. University of Exeter
To learn from experience and to unleash the creativity, a novelty approach has been proposed with the main aim of detecting the different leakages of a fictitious L-town, containing a total of thirty three different pressure sensors in the whole infrastructure.
To tackle this problem, a new solution has been presented based on the prediction of the mean night pressure for each pressure sensor located in the infrastructure. After the predictions have been generated, the solution compares them with the real values to highlight those cases when the pressure was meaningfully lower than expected, being able to detect leakages in the infrastructure in a reliable way.
The machine learning algorithm has proven to be very accurate with barely 0.0235% of relative error rate, making a very reliable base for performing leakage detection. Based on those predictions a total of 109 different leakages have been found distributed inside 31 different pressure sensors.
As a main conclusion we can infer that it is possible to build a leakage detection system based on the anomalies found in the pressure during nights, with the help of the stability of this attribute and the great accuracy of the machine learning algorithm the solution was able to make predictions with barely 0.0235% of relative error rate and being able to identify 102 different leakages