abstract
- © 2019 IEEE.In this paper, the one leak isolation problem in a water pipeline is tackled using a Time Delay Neural Network. This scheme comes as an alternative to achieve better computing performance since the classical model-based methods usually have high workloads due to the pipe mathematical model complexity compared with the leak dynamics speed. The Neural Network structure could have better time performance exploiting the parallel architecture of some electronics devices like an FPGA.The authors propose a scheme where, due to the difficulty in obtaining training data from a real pipeline, a mathematical model is used to generate synthetic training data. Such training data is obtained using different leak magnitudes and leak positions and it is also corrupted by random noise in order to emulate real data pipe. Finally, to show the potentiality of this method, some results are presented by using real-noisy databases coming from a pipeline prototype.Following the classical leak diagnosis hypothesis, only flow and pressure sensor at both ends of the aqueducts are used for the treatment.