Fusion-Inception-Aux: An Auxiliary-Supervised Multi-Scale Fusion Network for Distributed Acoustic Sensing Event Recognition

Abstract

Distributed acoustic sensing (DAS) event recognition is challenging because disturbance events may have similar temporal waveforms, heterogeneous channel responses, and strict false-alarm requirements in practical monitoring systems. This paper proposes <bold>Fusion-Inception-Aux</bold>, an auxiliary-supervised dual-branch framework for DAS event classification. The model uses inception-style temporal encoders to extract multi-scale disturbance features from two channel groups and applies auxiliary branch supervision to improve branch-wise representation learning. The final decision is obtained by adaptive aggregation of the branch-level and fused predictions. To avoid evaluation leakage, channel grouping is selected only from the training partition, and the main comparison is conducted under matched data splits, preprocessing settings, and training budgets. Experiments are conducted on a six-class, 12-channel Phi-OTDR dataset with 12,335 training samples and 3,084 test samples. Under five-seed evaluation, Fusion-Inception-Aux achieves <inline-formula><tex-math notation="LaTeX">$0.9964 \pm 0.0015$</tex-math></inline-formula> accuracy and <inline-formula><tex-math notation="LaTeX">$0.9964 \pm 0.0016$</tex-math></inline-formula> macro-F1, compared with <inline-formula><tex-math notation="LaTeX">$0.9936 \pm 0.0004$</tex-math></inline-formula> accuracy and <inline-formula><tex-math notation="LaTeX">$0.9934 \pm 0.0004$</tex-math></inline-formula> macro-F1 for the Fusion-CNN baseline. With the full training recipe, the proposed model reaches <inline-formula><tex-math notation="LaTeX">$0.9979 \pm 0.0004$</tex-math></inline-formula> accuracy and <inline-formula><tex-math notation="LaTeX">$0.9978 \pm 0.0005$</tex-math></inline-formula> macro-F1. Ablation results indicate that multi-scale Inception encoding and auxiliary supervision are the most important components, while the PatchX interaction module provides no consistent additional gain under the current setting. External validation on a three-class Mendeley DAS subset achieves <inline-formula><tex-math notation="LaTeX">$0.9844 \pm 0.0061$</tex-math></inline-formula> accuracy and <inline-formula><tex-math notation="LaTeX">$0.9844 \pm 0.0061$</tex-math></inline-formula> macro-F1, showing comparable performance on an independent public dataset. Complexity profiling further shows that Fusion-Inception-Aux improves recognition accuracy over Fusion-CNN with a moderate increase in model size and inference cost.