An Efficient Separation and Identification Algorithm for Mixed Threatening Events Applied in Fiber-Optic Distributed Acoustic Sensor

An accurate and fast recognition of some noticeable threatening events has been proven effective in fiber-optical distributed acoustic sensor (DAS). However, it is still challenging to find an efficient way to realize the accurate potential threats detection and identification. Especially in complicated environments, some weak threatening signals such as manual digging are usually submerged by the strong background noises, in which the influence of these interferences is unavoidable. Unfortunately, these effects of the mixed heavy interferences may cause ignored cases of the alert of potential threats, which even causes significant economic loss. In this work, an accurate and effective multisource signals separation and recognition algorithm is proposed to achieve the identification of the potential threats submerged in multisource noises for fiber optic DAS. First, the overlapping interferences in complicated environments can be effectively denoised by the proposed multisource signals separation algorithm. Then the multiscale features of different signal targets can be automatically extracted and identified by an attention-based multiscale convolution neural network (MS-CNN) model. In the field tests, four types of mixed multisource signals are performed to validate the effectiveness of the proposed algorithm. Finally, the field test results show that the recognition rate of the mixed signals is improved from 53.82% to 95.43% by the proposed algorithm. Besides, the performance of three network models based on the same database is compared. The final results prove that the attention-based MS-CNN model can obtain improved training speed and recognition accuracy, compared with the 1DCNN model and MS-CNN model. The proposed algorithm has an excellent performance for mixed threat identification in various complicated interferences.