Research On Signal Extraction And Pattern Recognition Methods Of Distributed Optical Fiber Vibration Sensing System

The main application form of distributed optical fiber vibration sensing system is Phase-Sensitive Optical Time-Domain Reflectometer,using the optical fiber as the sensor directly.With unique advantages of low-power and anti-electromagnetic interference,it can monitor all types of vibration in real time,and it has important application value in the fields of perimeter security,oil and gas pipeline leakage monitoring,urban pipeline and building structural health monitoring.The system still has some problems at present,such as inaccurate acquisition of vibration information,low signal-to-noise ratio,insufficient accuracy and efficiency of the recognition algorithm,therefore the research of vibration information's accurate extraction and effective pattern recognition algorithm are very important.The main research content of this thesis includes the quadrature phase demodulation method,used to achieve restoring vibration signal accurately in the ?-OTDR system.Variational mode decomposition and permutation entropy are used to remove the random noise of the system,and a wavenet deep learning algorithm is used to realize the pattern recognition of time-space vibration data.The results of simulation and experimental measurement are excellent.Finally,accurate and fast pattern recognition of different vibration types has been achieved,and it help the ?-OTDR system to realize its best performance in practical applications.Specific works performed and results achieved are as follows.1.Deduct the working process of heterodyne coherent ?-OTDR system detecting the vibration,and analyze polarization fading noise and coherent fading noise in the system.2.Adopt the quadrature I/Q phase demodulation method.The vibration source is simulated by piezoelectric ceramic for conducting experimental verification.the results show that the phase demodulation method can accurately demodulate the single-frequency disturbance in the range of 10Hz?1500Hz and mixed frequency disturbance,the disturbance amplitude and demodulated phase amplitude are positively correlated,and the linear fitting goodness is about 0.9879.3.The permutation entropy is applied to the Variational Mode Decomposition algorithm to determine the optimal decomposition layer K value,and realize the denoising function.The effect of the Empirical Mode Decomposition,Complete Ensemble Empirical Mode Decomposition and Modified Ensemble Empirical Mode Decomposition in denoising is compared and analyzed by simulation.The results show that,the VMD denoising algorithm has the best completeness and orthogonality,which are 0.1694 and 0.0030 respectively,and can effectively suppress the mode mixing and fake component of EMD method,The SNR is up to 17.0834 dB;and compared with CEEMD method and MEEMD method,the processing efficiency is even higher,the computing time is shorter,which is only 0.3331 s.4.Based on the temporality of input vibration signal,wavenet model is applied in ?-OTDR system for pattern recognition;its internal causal convolution structure can fully analyze the context relationship of input signals,and it can improve the accuracy of recognition.The dilated convolution structure further expands the model's perception range of input signals.The residual network makes the model converge faster and improves the training efficiency.From the perspective of model structure,the process,advantages and disadvantages of one-dimensional convolutional neural network and Long Short-Term Memory for time sequence input signal recognition are analyzed and compared.5.In the end,we built an experimental system to demonstrate the effectiveness of the approach.We collected six actions of walking,wheel rolling,electric drill disturbance,wind blowing,raining and net touching,after phase demodulation and VMD denoising,a time-space array of 50×50 was formed,data sets produced;they form the input to the one-dimensional CNN,LSTM and Wave Net models respectively for training and testing.The experimental results show that,compared with the one-dimensional CNN and LSTM,the recognition accuracy of Wavenet is as high as98.55%.And it consumes less training time,about 200 s.Also,its signal detection process takes only 30 ms,which can meet the real-time application requirement.The proposed pattern recognition method has high accuracy and good real-time performance.It should be of great significance for the application of ?-OTDR systems in perimeter defense,construction site protection and other practical scenarios.