Pipeline leakage poses a significant threat to critical infrastructure, including energy transportation and urban water supply systems. It creates serious safety hazards and results in substantial economic losses. While traditional detection methods like drone inspections, acoustic sensing, and infrared thermography are effective for regional monitoring, they face limitations in large-scale, complex environments, such as low efficiency, poor real-time performance, and limited accuracy. Distributed acoustic sensing (DAS) technology, which uses optical fibers to capture and analyze pipeline vibrations in real time, has emerged as a promising solution for large-scale monitoring. However, research on efficient DAS algorithms tailored for multi-scenario and multi-medium environments remains scarce, and existing algorithms require improvement in feature extraction and global relationship modeling. To address these challenges, this paper proposes a hybrid CNN-Transformer architecture designed to enhance the classification accuracy and efficiency of pipeline leakage signals across diverse scenarios.

The proposed CNN-Transformer model integrates the local feature extraction capabilities of CNNs with the global dependency modeling advantages of Transformers. DAS signals are first transformed into spectrograms via short-time Fourier transform (STFT) and processed by a spectrum feature extraction module. Concurrently, the signals are fed into a multi-scale feature extraction module, which includes a coarse-fine granularity submodule and a Transformer module. The submodule extracts fine-scale features (capturing local changes and high-frequency patterns) and coarse-scale features (representing broader trends and low-frequency information). These multi-scale features are then refined in the Transformer module to capture global dependencies and semantic relationships, enabling the model to deeply understand the intrinsic patterns of the signals. Finally, the combined feature vectors are passed into a classification module, where linear transformations further refine the features for accurate leakage identification.

Ablation experiments (Fig. 6) demonstrate that incorporating the Transformer module significantly enhances the model’s ability to capture global dependencies. Multi-layer Transformer integration deepens hierarchical feature extraction and boosts network representation capabilities. Comparative experiments with traditional models [Fig. 7(b)] show that the proposed model achieves an average accuracy of 97.466% across ten validation tests, outperforming classical models with minimal accuracy fluctuations, thus proving its robustness. Performance metrics analysis (Fig. 8) confirms excellent precision, recall, and F1-score results, validating the model’s structural effectiveness. In multi-scenario comparisons [Fig. 9(a)], the proposed model achieves an average false positive rate of 0.3% across seven industrial monitoring scenarios, a 42% improvement over baseline models. Although the Transformer module increases inference latency to 4.3 ms, this still meets industrial real-time requirements. t-SNE dimensionality reduction analysis (Fig. 10) reveals that post-extraction features form tightly clustered, well-separated categories in the reduced space, highlighting the model’s effectiveness in capturing critical signal distinctions.

The proposed CNN-Transformer architecture offers an efficient and reliable solution for DAS-based pipeline leakage monitoring. By combining spectrum feature extraction with global dependency modeling, the model achieves high classification accuracy across diverse scenarios (underground, underwater, aerial) and media types (liquid, gas). This study advances the field of DAS pipeline leakage signal recognition and provides new algorithmic insights. Future work will focus on expanding datasets to cover more leakage scenarios and integrating advanced algorithms to further enhance performance and applicability.