Pattern Recognition of Distributed Optical Fiber Vibration Sensors Based on Resnet 152

In recent years, traditional perimeter security system is gradually replaced by optical fiber distributed vibration sensing system, as it has superior advantages such as high sensitivity, fast response, and simple structure. However, it is still challenging to accurately realize multievent pattern recognition in practical applications. Accurate pattern recognition can reduce the false alarm rate and significantly increase the stability of the optical fiber system. In this article, we proposed a pattern recognition approach based on short-time Fourier transform (STFT) and Resnet 152-based neural network. First, the vibration signal containing high-frequency information was extracted through a median filter. Second, STFT was used to convert a 1-D time-domain signal to a 2-D time–frequency signal. The feature dimension of optical signals was expanded. Third, the redundant information would be removed by dividing the high-, medium-, and low-energy segments. Finally, the preprocessed optical signals were sent to Resnet 152 convolutional neural network (CNN) model for pattern recognition. To verify the effectiveness of the proposed scheme, field tests with nine sensing events (climbing, crashing, cutting, kicking, knocking hard, knocking lightly, no intrusion, pulling, and waggling) have been experimentally carried out. It is demonstrated that the average recognition accuracy of the nine common sensing events is 96.67%, and the detection time is 0.2391 s. The feasibility of deep CNN in solving pattern recognition has been proved.