Multi-Scale Frequency-Aware Transformer for Pipeline Leak Detection Using Acoustic Signals

Area of Science:

• Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Pipeline leak detection using acoustic signals faces challenges with time-frequency analysis, noise, and prior knowledge integration.
• Existing deep learning methods struggle to effectively utilize acoustic data for reliable leak identification.

Purpose of the Study:

• To propose a novel Multi-Scale Frequency-Aware Transformer (MSFAT) architecture for enhanced pipeline leak detection.
• To address limitations in current AI-based acoustic signal analysis for industrial measurements.

Main Methods:

• Developed MSFAT with a frequency-aware embedding layer for joint time-frequency feature learning.
• Incorporated a multi-head frequency attention mechanism and an adaptive noise filtering module.
• Utilized a multi-scale feature aggregation mechanism for robust global representation.

Main Results:

• MSFAT achieved 97.2% accuracy and a 10.9% improved F1-score compared to standard Transformers.
• Demonstrated robust performance across signal-to-noise ratios from 5 to 30 dB.
• Ablation studies confirmed the significant contribution of frequency-aware mechanisms.

Conclusions:

• The MSFAT architecture effectively integrates domain-specific knowledge into AI for superior pipeline leak detection.
• The proposed method offers enhanced precision, reliability, and adaptability in complex industrial environments.