Intrusion Identification Method for Φ-OTDR Systems Based on Few-Shot and Multi-Task Learning

Aiming at the problem of reduced recognition accuracy caused by insufficient number of samples when deep learning algorithms identify multiple intrusion information on fiber optic fences in the perimeter security field, this paper proposes a phase-sensitive light time domain classification method based on model-independent meta-learning (MAML) and multi-task learning. Complete the classification and judgment of multiple tasks by correlating the information of primary and secondary tasks. After the MAML algorithm is trained to obtain model parameters suitable for the current sample distribution, it is then fine-tuned to quickly adapt to the new environment. Experimental results show that using only a simple convolutional neural network and a small number of samples for each category, the recognition accuracy of the intrusion action area can reach 98.73% when the verification set of the 6-classification task can fully identify the intrusion action. This method can simultaneously complete two different identification tasks with a small number of samples to achieve event identification and regional location of perimeter intrusion signals.