Using neural networks to monitor piping systems

Abstract The paper proposes a state estimation technique, which uses Artificial Neural Networks (ANN) to monitor the status of piping networks carrying hazardous fluids, in order to identify and locate spills and leakages. A Multilayer Perceptron ANN is used to process pressure and flow rate information coming from a limited number of sensors distributed across the network. The ANN is trained on different sets of input data, which characterize several states of the fluid network under normal and abnormal operating conditions. During the running phase, it acts as a classifier in order to estimate the actual system status and pinpoint leaks, based on available information, thereby solving the stated inverse problem. A two‐level architecture is selected, composed of a main ANN at the first level, to identify the branch in which the leakage occurs, and several branch‐specific ANNs at the second‐level to accurately estimate the magnitude and location of the leaks. After describing the proposed methodology and the system architecture, we present a realistic application example in order to show the technique's potential.