A Fast recognition method for DVS perimeter security intrusion signals based on transfer learning

Abstract

In the perimeter security intrusion signal monitoring based on the Distributed Optical Fiber Vibration Sensing (DVS) system, the common approach for identifying intrusion signals in new scenarios is to collect large amounts of data and perform large-scale, long-term retraining of the initial network. However, this process consumes a significant amount of human resources and computational power each time. To address this issue, this paper proposes a transfer learning method. Using 53,000 sample data from six types of events collected in the laboratory, an SE-Densenet model was trained for 300 epochs on three NVIDIA A100 Tensor Core GPUs at the National Supercomputing Center in Jinan. Then, through the transfer learning approach, fine-tuning was performed on the new scenario by freezing 1/2 and 1/4 of the network layers and applying a feature extractor method. The fine-tuning training was carried out on a laptop equipped with 16 GB of memory and an Intel Core i7-12700 CPU. Experimental results show that, after 19.81 min of retraining using 600 data points from the new scenario, the feature extractor method achieved a test accuracy of 97.53 %. Compared to the other two methods, the training time was reduced by 20.87 min and 9.34 min, respectively, while the test accuracy increased by 13.76 % and 3.83 %. Compared to retraining with 6000 sufficient data, the transfer learning method improved training time by 6 to 13 times while achieving similar recognition accuracy. This study provides a new, convenient, and efficient method for perimeter security intrusion signal recognition, enabling quick adaptation to new scenarios and application in common engineering upper-level computer systems.