Multi-feature Fusion Network for Classification of Pipeline Magnetic Leakage Signals

Abstract

Purposes Leakage detection in underground pipelines such as prestressed concrete cylinder pipes is a crucial aspect of urban infrastructure management and maintenance. In this study, an innovative magnetic anomaly multi-feature fusion network (MMF) is designed and introduced to identify weak magnetic distribution types in underground pipelines. Methods The network leverages standard orthogonal basis functions (OBF) and minimum entropy detection (MED) features were used to comprehensively and accurately capture the complex characteristics of magnetic leakage signals. First, magnetic anomaly detection was conducted by using OBF and MED at different radial distances to acquire measured target magnetic field features. Second, an MMF was devised to integrate magnetic field features, and a multi-head attention mechanism was incorporated to capture intricate relationships and features within the sequence of magnetic fields. Finally, a multi-feature entropy weighting method was employed to allocate network weights on the basis of input feature entropy. Findings Experimental results demonstrate that the MMF network achieves a precision of 98.86% in anomaly classification, with an AUC evaluation result of 99.25%. Additionally, the model is more streamlined, exhibiting higher computational efficiency, and is capable of delivering satisfactory performance within a relatively short training period.

Keywords

• signal detection and classification
• multi-feature fusion network
• entropy weighting method
• multi-head attention mechanism