Research On Pattern Recognition Method Of Distributed Optical Fiber Sensing System Based On Φ-OTDR For Perimeter Security

Based on phase-sensitive optical time domain reflectometer(Φ-OTDR)distributed optical fiber sensor technology can accurately detect and locate multiple disturbances at the same time,it can be used for long-distance real-time monitoring and accurate positioning.It has been widely used in engineering structure health monitoring,optical perimeter protection and oil and gas pipeline safety protection system and other scenarios.However,in actual monitoring system of Φ-OTDR,the external natural environment interference is easy to introduce background noise in the process of data acquisition,resulting in the sample pollution of human intrusion events,reducing the recognition accuracy.In order to reduce false alarms in the sensing system and improve the event recognition rate of Φ-OTDR in complex environment,in this paper a sample feature correction algorithm based on sample feature weighting is proposed.By establishing an accuracy evaluation method and a weight allocation scheme based on sample feature correlation,the method combined with BP(Back Propagation)neural network can accurately identify disturbance events.Experiments have verified that this algorithm can effectively improve the distinguishability of features,thereby improving the recognition rate of various disturbance events.In the signal preprocessing and feature extraction stage,the signals are processed by normalization and data partitioning.After that,36 features were extracted from time-domain and frequency-domain.The impact of the number of features on the recognition accuracy is analyzed through the Fisher features selection method.In the aspect of feature optimization algorithm,in this paper a feature correction method based on sample feature weighting is proposed.The feature correction algorithm was used to optimize the noise sample features in the dataset,improving the sample feature discrimination.In terms of disturbance event recognition algorithm,in this paper BP neural network is used as a classifier,combined with feature correction algorithm to construct SW-BP(Sample weighting BP)pattern recognition method.A disturbance signal recognition model of Φ-OTDR has been obtained through training.In the experimental verification and analysis stage,a disturbance event recognition experiment has been established in a real environment.By establishing the distributed fiber optic sensing system of Φ-OTDR,signals from four disturbance events have been collected and a dataset has been established.Then the BP neural network method and the SW-BP method are used to train and verify the data.The SW-BP method is used for identification test.The average recognition accuracy of four disturbance events is 99.38%,The SW-BP method can improve the recognition accuracy by 3% compared to the BP neural network method,and the false alarm rate is also significantly reduced.A generalization test was conducted on the trained model by creating a new test dataset in an experimental scenario.The experimental results show that the SW-BP method has better generalization ability.The anti-noise test of the two methods was conducted by artificially mixing noise samples into the training samples.The SW-BP method can significantly improve the generalization accuracy by 4% compared to the BP neural network method.The SW-BP method has a significant improvement in average accuracy even in the presence of noise interference.The method proposed in this paper has more advantages in recognition accuracy,generalization,and anti-interference,and can effectively improve the reliability and robustness of Φ-OTDR in complex environments.