Research On High Performance Pattern Recognition Algorithm In Distributed Fiber Acoustic Sensing System

The Chinese government has exerted great efforts to develop the large-scale infrastructure such as bridges,highways,railways,rail transit,power lines and natural gas pipelines in recent years.During the ‘service’ process that lasts for decades or even hundreds of years,the structure of the large-scale infrastructure will inevitably suffer damage accumulation and resistance attenuation due to the adverse effects of environmental corrosion,material aging,long-term loading,natural disasters and manmade damage.Once a structure is damaged,it will lead to a sudden catastrophic accident.The casualties,economic losses and social impact caused by the accident are very large.Therefore,there is an urgent need to establish a health monitoring system for large-scale infrastructure structures,grasp the health status of infrastructure structures in real time,find the location of structural damage in time and accurately identify the cause of damage,so as to send out early warning signals and take corresponding measures to avoid unnecessary waste of resources to ensure the safe and reliable operation of infrastructure.As an important branch of distributed optical fiber sensing technology(DOFS),Distributed Optical Fiber Acoustic Sensing technology(DAS)can realize long-distance,distributed and real-time quantitative monitoring of dynamic strain along the fiber,which has a very broad application prospect in the field of large-scale infrastructure structure health monitoring.Although DAS technology can realize the measurement of disturbance signals,to solve specific engineering application problems,it is necessary to carry out pattern recognition on the events represented by the signals to distinguish valid and invalid information.At present,there are still the following problems in the pattern recognition of DAS disturbance signals: First,the problem of coherent fading.DAS uses a narrow linewidth laser to cause interference between the back Rayleigh scattered light(RBS)in the pulse range to obtain high sensitivity,but it also inevitably brings about the problem of coherent fading.Due to the coherent fading,the RBS intensity at certain positions of the fiber approaches zero,which prevents the sensor from detecting these areas,which forms fading dead zones.When the signal drops into the fading dead zone,it may be submerged in noise,and the signal-to-noise ratio will deteriorate sharply,resulting in a large error in the phase demodulation result,and in serious cases,it will cause false alarms,which greatly limits the performance of DAS system.The second problem is feature extraction,which is the core step of pattern recognition.Whether the extracted features are representative will directly affect the classification results of classifier.The existing feature extraction methods have problems that the feature expression is too simple,unrepresentative and does not consider the amount of feature data,which make low reliability of recognition in complex environment.In response to the above problems,a coherent fading suppression method has been proposed in this paper,which can reconstruct the vibration signal with high fidelity under the premise of using only ordinary single-mode sensing fiber without changing the structure of the traditional DAS system.The results show that the phase discrimination accuracy of the coherent fading suppression method is improved by 19.99% compared with the traditional phase discrimination method.Then,we used the DAS equipment developed by our research group to conduct large-scale infrastructure external field tests.When performing feature extraction on the 5 kinds of disturbance signals,four features of energy,average value,square difference and disturbance duration are extracted in the time domain,and in the frequency domain,we propose to perform binary processing on the short-time Fourier transform(STFT)spectrum,extract 5 features of area,perimeter,compactness,number of connected domains and Euler number of binary image and use SVM classifier and feedforward neural network to classify feature vectors.The recognition accuracy rates of the 5 types of disturbance events are 97% and 98%,respectively.Experimental results show that the method proposed in this paper can greatly improve the phase discrimination accuracy of DAS system and the recognition accuracy of disturbance events in practical engineering applications,which will play a good role in promoting the application of DAS equipment in health monitoring of large infrastructure structures.