Distributed fiber optic acoustic sensing system intrusion full event recognition based on 1-D MFEWnet

Abstract

Proper detection of the full range of intrusion events is of paramount significance to distributed fiber optic sensing perimeter security systems. Traditional neural networks for intrusion event recognition are constrained by the training dataset, that is, they cannot detect intrusions outside of the training dataset. However, in real complex environments, the dataset by manually obtained is far fall short of encompassing all possible real-world data. This limitation can lead to inaccuracies of identification in the distributed fiber optic sensing system not being able to identify correctly, which causes immeasurable losses. In order to address the aforementioned issues, this paper presents a 1D MFEWnet model, which completes the effective differentiation of all datasets by means of a Multi-Feature branch 1-dimensional Convolution Neural Network, followed by fitting the activation vectors after the recognition of known datasets to a Weibull distribution, through the improved Euclidean distance tracing algorithm. This approach allows for the extraction and identification of additional intrusion signals while providing the ability to recognize and reject unknown interference events. In the experiments, a distributed fiber optic sensing system was established to collect event signals. For three known event categories, the highest recognition accuracy is up to 99.6%. After adding 2 unknown event categories randomly, the accuracy remained at a commendable 96.9%. This innovative methodology ensures the accuracy of target recognition under the introduction of all conceivable events and improves the robustness of the distributed fiber optic perimeter security system.