Multisource Event Classification in ϕ-OTDR Through Sample Feature Synthesis

Traditional deep learning methods typically deal with single-source events in $\Phi $ -optical time domain reflectometry (OTDR) event recognition, and the classifier training requires a significant amount of data that must comprehensively cover all expected event types. Given the enormous combinations of multisource events, which consist of several single-source events, it is impractical to obtain all types of multisource events in practice when the single-source events increase. Driven by the concept of feature fusion, we implemented a mixed class and reinforcement learning guided training method (MC-RLGTM) to synthesize multisource event features from single-source event samples, thereby circumventing the substantial workload of collecting actual data and avoiding the immense effort involved in data collection. With the synthesized features of multisource event, the final classifier can be then trained in a supervised way. The experimental validation demonstrates that the method can achieve 66% accuracy for multisource events in a seven-event classification task, with a 45% improvement compared to the traditional classification method.