SP-MSResNet: A Multiscale Residual Network With Strip Pooling Module for Intrusion Pattern Recognition

The perimeter security system based on distributed optical fiber sensing is of great significance for the monitoring and security of military and key industrial areas. Most of the existing recognition methods for distributed optical fiber disturbance signal rely on artificial feature extraction method based on time domain or frequency domain, combined with decision tree or support vector machine (SVM) to perform the recognition operation. However, these methods usually have the problem of long recognition time and high error rate, which brings great obstacles to the practical application of perimeter security system. To solve the above problems, this article proposes a multiscale residual network with strip pooling (SP-MSResNet) module, which takes into account the advantages of traditional signal frequency-domain analysis methods, and the powerful feature extraction capability of deep network model. First, the original 1-D intrusion disturbance signal is decomposed into five levels in the frequency domain by discrete wavelet transform (DWT), and the obtained five approximate horizontal coefficients are connected in parallel with the original signal to form a new 2-D data. Then, the 2-D data were applied to the SP-MSResNet for the training of recognition task. The experimental results show that the average recognition accuracy reaches 99.17% on the dataset of six types of intrusion disturbance signals (wind blowing, light rain, heavy rain, knocking, impacting, and slapping), which has more advantages compared to other methods mentioned in the literature.