Lightweight event recognition framework for Φ-OTDR systems with space-frequency fusion based on self-exclusion relational knowledge distillation

Phase-sensitive optical time-domain reflectometry (Φ-OTDR) systems are widely utilized in large-scale perimeter security and critical infrastructure protection. However, disturbance signal classification in Φ-OTDR systems heavily relies on complex model structures. Achieving high classification accuracy often requires high computational demands and substantial costs. This paper proposes an ultra-lightweight classification framework named space-frequency self-exclusion relational knowledge distillation (SF-SERKD). This framework constructs a three-channel feature map by using the short-time Fourier transform (STFT) combined with gradient analysis to capture the frequency-domain and spatial mutation features of signals. In addition, the framework introduces an innovative self-exclusion relational knowledge distillation (SERKD) strategy, which explicitly excludes the target sample’s own teacher features during the nearest-neighbor matching process. This strategy forces the student model to abandon feature alignment with high-noise samples or isolated samples, thereby reconstructing a robust and shared intra-class feature structure. Validation results in two different scenarios show that the student model achieves classification accuracies of 99.81% and 99.29%, alongside ultra-low nuisance alarm rates of 0.0004 and 0.0037, respectively, while only requiring 0.007M parameters. Notably, the accuracy of the student model in the second scenario is 1.87% higher than that of the teacher network. These results elucidate the effectiveness and generalization of the proposed method for real-time vibration event recognition in Φ-OTDR systems.