A Noise-Robust CNN-KAN Architecture with Dual Attention for Enhanced Event Identification in Φ-OTDR Measurement Systems

Phase-sensitive optical time-domain reflectometry (Φ-OTDR) is a well-established technique for the distributed measurement of dynamic strain along the optical fiber. However, the metrological reliability of event identification is inherently degraded by coherent fading noise, laser phase fluctuations, and environmental interference, which corrupt the acquired backscattering signals and limit the measurement accuracy of the sensing system. This paper presents a novel signal processing architecture that enhances the information extraction capability within the Φ-OTDR measurement chain. By integrating multi-scale residual convolutional feature extraction with dual channel-spatial attention mechanisms and an improved Kolmogorov-Arnold Network (KAN) classifier employing learnable radial basis function splines, our approach robustly suppresses measurement noise to improve the fidelity of extracted event signatures. The hybrid architecture addresses the limitations of conventional threshold-based detection methods that suffer from poor estimation accuracy under low signal-to-noise ratio conditions. Experimental evaluation on the BJTU dataset demonstrates a significant improvement in measurement precision, achieving 99.87% classification accuracy with 6.2 ms end-to-end inference latency and 161 samples/s throughput—representing a 2.4× speedup over Φ-GLMAE and eliminating the 2.5 ms STFT preprocessing overhead of STFT-AECNN, while maintaining real-time suitability for embedded deployment. Ablation studies quantitatively validate the contribution of each component to noise robustness and measurement reliability, demonstrating that dual attention mechanisms provide the largest single accuracy gain (0.46%), while the KAN classifier and RBF splines collectively enable 79% error reduction versus CNN baselines. This work offers an effective solution for high-fidelity distributed acoustic measurement in challenging operational environments.