Research On Perimeter Intrusion Detection Based On Deep Learning And Feature Fusion

The perimeter security system based on distributed optical fiber sensor is an intelligent detection system used in perimeter security field.Due to the advantages of high precision and sensitivity,good anti electromagnetic interference and corrosion resistance,distributed optical fiber sensor can be used for long-distance detection.On the basis of it,combined with the deep learning technology with stronger robustness,it has great research value for the study of perimeter security with complex environment.The purpose of this paper is to improve the recognition rate of disturbance signals in the perimeter security system based on distributed optical fiber sensors.Therefore,a method of signal classification and detection based on deep learning and feature fusion is proposed.Firstly,the signal is represented by time-frequency to break through the limitation of one-dimensional signal analysis,and then the time-frequency and pseudo color dynamic spectrum of the signal obtained by timefrequency processing are used as the input of neural network for model training.Through the improved MobileNet training in this paper,the depth characteristics of the signal are extracted,which are combined with the time-frequency characteristics of the signal extracted by traditional methods The fiber optic disturbance signals are classified and detected.The main contents and work of this paper are as follows:(1)This paper analyzes the importance of perimeter security system,system components,composition and research value,introduces the functional characteristics of different distributed optical fiber sensors,and compares the characteristics of several different optical fiber sensors.The Rayleigh scattering distributed optical fiber sensor is selected as the basic sensor of this paper,and the perimeter security system architecture for perimeter monitoring based on the sensor is designed.(2)Before the disturbance event detection,firstly,the disturbance signal collected by the optical fiber sensor is processed by wavelet de-noising,and a new threshold function of wavelet denoising is proposed to effectively improve the effect of noise reduction.Then,the time-frequency manual feature extraction is carried out for the noise reduced disturbance signal,and SVM is used as the disturbance signal classifier for comparative test.(3)The disturbance signal is extracted and constructed from the feature image.As one-dimensional vibration signal,disturbance signal can only be expressed in time domain.In this paper,the time-frequency characteristics of disturbance signal are processed to break through the limitation of feature extraction in one-dimensional time-frequency analysis.Firstly,the de-noising signal is decomposed into seven layers of wavelet packets,and the energy corresponding to the wavelet packet transform coefficients of each subband is calculated,and the time-frequency diagram of the signal is drawn.At the same time,the fast Fourier transform is used to process the original one-dimensional signal,and the pseudo color dynamic spectrum of the signal is constructed as the input of the network to get the depth characteristics of the signal.(4)Based on the MobileNet network,the feature image is used as the input of the Inception-V3 network,the Xception network,the MobileNet network model and the improved MobileNet network model,and the network is trained under condition of small samples to obtain the depth characteristics of the signal.Finally,the depth feature and the time-frequency feature(that is,the time-domain feature)are used to get the depth feature of the signal And wavelet packet energy feature)to classify and recognize the disturbed signals.(5)In this paper,five kinds of disturbance events are used to verify the correlation experiments.Traditional machine learning and deep learning combined with feature fusion method are applied to the same data set.The experimental results show that the pseudo color dynamic spectrum is used as the input of the improved MobileNet network model,and the depth feature of the output of the bottleneck layer and the time-frequency feature of the signal are fused to process the feature fusion,which can more effectively and accurately detect the disturbed signal.The average accuracy of the algorithm can reach 99.45%.This paper verifies the effectiveness and efficiency of the method through comparative experiments Reliability.