The Research Of Signal Recognition For Fbg Signals In Perimeter Monitoring Based On CNNs

The perimeter security system based on optical fiber sensing technology is gradually applied to various fields of society due to its advantages of corrosion resistance,anti-electromagnetic interference,layout flexibility and high stability.Traditional signal recognition algorithms based on signal processing and feature engineering must undergo cumbersome feature extraction and selection steps.These work require too much professional experience and cannot meet the design requirements of perimeter security system with increasingly complex environment.On the other hand,the actual engineering environment of perimeter security is susceptible to noise interference and device load.These non-human factors interfere with the normal use of the system and often cause false alarms.In view of the above problems,this paper proposes a convolutional neural network intrusion signal recognition method based on the raw data signal.The main research work of this paper is as follows:(1)The grating array perimeter security system based on fiber-optic sensing technology is constructed.Combined with the working mechanism of fiber-optic sensing,the acquisition and pre-processing methods of intrusion vibration signals are given.Then,the method of signal sample segmentation is designed according to the characteristics of the acquired signals.According to the monitoring requirements of security system and the characteristics of human action,the intrusion behavior of the perimeter security system is divided.Based on the intrusion behavior,the intrusion signal sample database is established.In consideration of the practical application environment,the sample database has differentiated the railings with different degrees of damage for behaviors.(2)Aiming at the cumbersome operation and inefficiency of traditional feature engineering signal recognition methods,an OSCNN signal recognition method acting on the original time series signal is proposed.This method overcomes the defect of partial vibration information loss caused by artificial feature extraction,and it uses the network layer of the OSCNN model to implement autonomous feature learning.The OSCNN model is composed of the basic network layer like other convolutional neural network models.Its biggest feature is that the convolutional layer has structure of the large convolution kernel in the first layer and the small convolution kernel in the latter layers.To verify the advantage of OSCNN with other signal identification method,OSCNN is compared with traditional signal classification model and classical convolutional neural network model in the experiment.The classification results of several methods are analyzed and compared under different signal classification scenarios.Finally,the experimental results show that OSCNN not only ensures time efficiency,but also greatly improves the classification accuracy rate.(3)In order to solve the problem that the classification effect is reduced due to the noise interference and the damaged of the railing,the A-OSCNN model with domain adaptive ability is proposed based on the OSCNN model.Combining the generalization ability of the Adam parameter optimization algorithm and the BN layer,the Adam-BN layer is added to the convolutional layer and the fully connected layer of the model.The experimental results also show that the improved A-OSCNN model has good adaptive performance.In order to verify the effectiveness of the improved A-OSCNN model in complex signal recognition environment,the model is applied to the signal recognition scene of noise interference and damage load,and OSCNN and other classification models are added to observe the anti-noise and self-adaptive ability of several methods.The final experimental results show that the A-OSCNN model has good anti-noise and adaptive performance.Compared with the traditional signal recognition method,the classification accuracy rate is greatly improved.And the improved A-OSCNN model has better adaptive ability than OSCNN and classical convolutional neural network model.