Deep-learning-enabled distributed acoustic monitoring of bird activity on overhead transmission lines

Abstract

To address the mismatch between the urgent need for bird-hazard prevention on transmission lines and the spatiotemporal limitations of existing monitoring approaches, and to support operation and maintenance (O&M) units in implementing targeted mitigation strategies, we propose a distributed acoustic sensing-based method for bird-call event detection and bird-call classification on overhead transmission lines. To resolve the difficulty of detecting bird calls under complex noisy conditions, we develop a bird-call event detection model that integrates time-frequency feature extraction with time-frequency joint attention. Using log-Mel spectrograms as input, the model achieves a detection accuracy of 97.48 ± 0.12%, with a detection latency of 7.3 ms for a single 0.5 s window. The detected signals are then further transformed into MFCC heatmaps, from which time-cepstral features are extracted using a multi-scale convolutional backbone. Combined with denoising autoencoder (DAE)-based feature enhancement, this framework enables bird-call classification under severe environmental noise. Ablation experiments show that each module contributes substantially to the overall classification performance. Across five independent runs on the fixed independent test set, the full classification model achieved a mean accuracy of 95.67 ± 0.16%, outperforming the adapted TSAM-Net by 5.61% points. These results demonstrate the feasibility, accuracy and robustness of the proposed method under simulated transmission-line conditions with complex noise, and provide a promising technical basis for future monitoring of bird activity along long-distance transmission lines.