A Contrastive Representation Learning Method for Event Classification in Φ-OTDR Systems

Area of Science:

• Fiber optic sensing
• Signal processing
• Machine learning

Background:

• Phase-sensitive optical time-domain reflectometry (Φ-OTDR) is vital for distributed acoustic sensing.
• Accurate event classification is essential for Φ-OTDR system deployment.
• Existing methods require extensive labeled data, hindering practical application.

Purpose of the Study:

• To introduce CLWTNet, a novel method for event classification in Φ-OTDR systems.
• To address the bottleneck of labeled data dependency in current methods.
• To develop a cost-effective and efficient solution for Φ-OTDR data analysis.

Main Methods:

• CLWTNet utilizes contrastive representation learning on unlabeled Φ-OTDR data.
• Time-domain signals are transformed into Short-Time Fourier Transform (STFT) images.
• Wavelet transform convolution is integrated to capture complex signal features.

Main Results:

• CLWTNet achieves competitive performance compared to supervised methods.
• The proposed method outperforms existing unsupervised methods.
• CLWTNet effectively extracts discriminative representations from unlabeled data.

Conclusions:

• CLWTNet demonstrates the efficacy of unsupervised representation learning for Φ-OTDR event classification.
• The method significantly reduces the need for costly data labeling.
• CLWTNet offers a practical and efficient solution for real-world Φ-OTDR applications.