Objective Distributed fiber-optic vibration sensing (DFOS) must simultaneously address two open-world challenges in long-distance security monitoring: "unknown-event intrusion" and "extreme imbalance among known classes." Traditional supervised models can only perform closed-set classification and label any disturbance absent from the training set as a generic "anomaly," failing to distinguish risk levels among events such as manual digging, rain washing, or cable cutting. This leads to high false-alarm and missed-alarm rates in the field. This paper aims to build a zero-shot deep-learning framework that combines "high-precision known-class classification" with "fine-grained unknown-class recognition," enabling DFOS systems to automatically discriminate among multiple potential threats without retraining, and providing reliable monitoring for pipelines, railways, borders, and other complex scenarios. Methods The framework consists of three core modules. 1) Multi-scale feature extraction: parallel 1-D CNN branches with small, medium, and large kernels capture fine, medium, and coarse-grained local features from phi-OTDR back-scattered spectra; each branch is followed by ReLU and max-pooling, after which a feature-pyramid network fuses multi-scale information into a unified representation that retains both high-frequency transients and low-frequency trends. 2) Bagging ensemble classification: to counter on-site class imbalance, independent CNN base classifiers are trained, and their posterior probabilities are averaged at inference, markedly reducing environmental noise and overfitting risks for rare classes. 3) Zero-shot prototype network: class-prototype vectors are constructed from the feature means of known classes in the embedding space; for unknown events, Euclidean distances to these prototypes are computed. An adaptive distance threshold implements a dual decision-below the threshold the sample is assigned to the nearest known class, above it is labeled unknown, and further distance ranking provides sub-class ordering of unknown types, achieving "open-set + fine classification." The entire network is trained end-to-end with a composite loss that weights cross-entropy and prototype contrastive loss. The optimizer is Adam with an initial learning rate of 0.001; if validation loss does not improve significantly within 20 epochs, the learning rate is multiplied by 0.75 until a minimum of 0.00001 is reached, and an early-stopping mechanism prevents overfitting. Results and Discussions Experiments are conducted on a real-world dataset collected along a 12 km buried pipeline using our self-developed phi-OTDR system. Four classes-"ambient, car, excavation, pedestrian"-are designated as known, and three additional unknown classes are introduced, maintaining significant time-frequency differences. Without using any labels from the unknown classes, the proposed framework achieves an overall accuracy of 97.6 % on known events and an average accuracy of 58.33 % on the three unknown events. Conclusions The proposed "multi-scale CNN-Bagging ensemble-prototype network" hybrid framework realizes, for the first time in the DFOS domain, zero-shot event recognition. It maintains approximately 98 % precision on known threats and distinguishes multiple unknown disturbance types without retraining, offering operators dual information-"risk level + event category. " This work overcomes the limitations of traditional closed-set recognition, allowing a single sensing system to adapt to open, dynamic, and imbalanced real-world scenarios, markedly reducing on-site false alarms and maintenance costs. Future efforts will introduce incremental learning and cross-domain transfer to shorten adaptation cycles for unknown classes and explore extensions to multimodal sensing networks such as acoustic emission and MEMS.
更多
原文链接
分享
引用
