Multiscale entropy and supervised manifold mapping for leak diagnosis in pipeline structural health monitoring

Abstract

To establish a lightweight yet highly discriminative diagnostic framework for reliable leak state identification using minimal signal data without relying on complex deep learning architectures, this paper proposes a novel approach integrating multiscale information entropy (MIE) and angle-distance-based supervised manifold mapping (SMM). The primary objective is to enhance the separability of high-dimensional acoustic features, particularly for short-duration signals. The method operates in two stages: MIE extracts entropy features across multiple time scales to capture cross-scale signal complexity, while SMM collaboratively optimizes intra-class compactness and inter-class separability, projecting features into a low-dimensional space with maximized discriminative power. This synergy directly addresses the common issues of insufficient cross-scale information capture and excessive model complexity in existing approaches. To validate the method, a pipeline leak experiment was conducted under no-leak (0 mm) and various leak aperture (1–6 mm) conditions. Using a Bayesian classifier with signal durations as short as 1 s, the proposed method was evaluated against multiple dimensionality reduction techniques. Results demonstrate clear separation of leak and non-leak samples in the three-dimensional SMM space, enabling intuitive visualization-based diagnosis without complex classifiers. The method maintains stable diagnostic performance even with 1-s signals, proving its ability to extract meaningful features from limited data, and achieves average recognition accuracy consistently exceeding 90% across all tested conditions—outperforming alternative methods. This study provides an efficient, interpretable, and high-performance solution for pipeline leak diagnosis, particularly well-suited for applications requiring reliable diagnosis from very short signal segments in data-limited scenarios.