Pipeline Leak Detection Using Acoustic Emission and State Estimate in Feature Space

Abstract

Existing acoustic emission (AE) signal-based methods for pipeline leak detection (LD) usually denoise the raw signals directly in signal space, then extract signatures from denoised signals, and finally classify normal/leaky states via classifiers trained using offline datasets. Their complex computational structures may limit their real-time application, especially, when they will be required to analyze massive amounts of data. Furthermore, these methods may not be effective in LD in real pipelines, where AE signals might be prone to constant fluctuation. This article proposes a novel technique to mitigate these issues. It combines a Kalman filter and an outlier removal technique to estimate the true state in feature space and identifies a leak through normalized distance from an unknown class to a well-known class with a threshold. The experimental results show that the proposed method achieves an average true detection rate (TDR) of 96.9% and an average omission rate (AOR) of 3.6% compared to existing methods, which achieve a maximum average TDR of 92% and a minimum AOR of 8.8%. Moreover, the proposed method can achieve these results in real time.