Research On Intrusion Signal Recognition And Classification Technology Of Optical Fiber Perimeter Security System Based On Machine Learning

With the development of the country and the society and economy,the demand for security in various fields has become increasingly strong.The stable operation of infrastructure such as water conservancy,electricity,oil and gas,minerals,communications,networks,finance,and national defense requires a reliable security system.The optical fiber is made of sufficient raw materials,and has stable and excellent physical and chemical properties such as light weight and corrosion resistance.The optical fiber security system based on the principle of optical fiber sensing has attracted the attention of the market and researchers.In order to improve the accuracy of the system's classification of intrusions,the classification and recognition of different event types combined with machine learning algorithms has become the mainstream research direction,and good classification results have been achieved.However,in most of the research work of algorithm recognition and classification,the coupling degree between the algorithm and the feature engineering relying on professionals is too high,which makes the security system need professional debugging in different scenarios,and the adaptability is not strong.In response to the above background and problems,this article has done the following work.(1)Comparative analysis of the performance and operation mechanism of three phase modulated optical fiber sensor,the Michelson interferometer is selected based on the actual situation of engineering.(2)Analyze the difference between the time-domain waveform and the frequency-domain amplitude spectrum characteristics of the collected five signal data.Through normalization,time-frequency transformation,two-dimensional folding,visualization and other data processing methods,five different input algorithms are determined for one-dimensional time-domain raw data,time-domain digital matrix,frequency-domain digital matrix,time-domain waveform,frequency-domain amplitude spectrum,etc.The data form of the model is used to support the subsequent algorithm verification work.(3)The detailed principles of support vector machines and artificial neural networks are introduced.CNN adheres to the idea of deep learning,the quantity and quality of features obtained by the network model learning are better than those obtained by manual design.Therefore,sending raw data that has only been simply processed into the convolutional network for recognition and classification can also obtain better results.Convolutional neural network is used for signal recognition,which reduces the coupling between artificial feature extraction and algorithm recognition and classification.(4)A one-dimensional convolutional network model is designed and constructed.Drawing lessons from the design ideas of residual convolutional networks,two two-dimensional simple residual network models are designed and constructed.Experiments verify the reliability and effectiveness of the simplified data processing method and the selection of CNN for signal recognition and classification.