Improved parallel convolutional neural network model for anomaly detection in fluid transportation system

Abstract

The presence of water leakages in distribution systems leads to issues such as water loss, and damage to infrastructure. It is very challenging to precisely identify leaks in distribution systems because leaks are unwanted happenings in water distribution networks (WDNs). Implementing supervised machine learning techniques requires precisely labeled data. Lack of labeled data makes it difficult to apply supervised machine learning techniques. This study suggests an innovative Enhanced Parallel Convolutional Neural Network and GhostNet (EPCNN-GhNet) for identifying leaks in fluid transportation systems, which consists of two crucial phases: normalization and anomaly detection. Initially, the input data is normalized using Min-Max normalization. An Enhanced Focal Loss function-based GhostNet (EFL-GhNet) and a Modified Batch Normalization-based Parallel Convolutional Neural Network (MBN-PCNN) are combined in a hybrid approach to detect abnormalities. The outputs of both approaches are combined using soft voting to determine whether a leak exists in the system.