Optical Fiber Vibration Signal Recognition Based on the Fusion of Multi-Scale Features

Xinrong Ma,Jiaqing Mo,Jiangwei Zhang,Jincheng Huang
DOI: https://doi.org/10.3390/s22166012
IF: 3.9
2022-01-01
Sensors
Abstract:Because of the problem of low recognition accuracy in the recognition of intrusion vibration events by the distributed Sagnac type optical fiber sensing system, this paper combines the traditional optical fiber vibration signal recognition idea and the characteristics of automatic feature extraction by a convolutional neural network (CNN) to construct a new endpoint detection algorithm and a method of fusing multiple-scale features CNN to recognize fiber vibration signals. Firstly, a new endpoint detection algorithm combining spectral centroid and energy spectral entropy product is used to detect the vibration part of the original signal, which is used to improve the detection effect of endpoint detection. Then, CNNs of different scales are used to extract the multi-level and multi-scale features of the signal. Aiming at the problem of information loss in the pooling process, a new method of combining differential pooling features is used. Finally, a multi-layer perceptron (MLP) is used to recognize the extracted features. Experiments show that the method has an average recognition accuracy rate of 98.75% for the four types of vibration signals. Compared with traditional EMD and VMD pattern recognition and 1D-CNN methods, the accuracy of the optical fiber vibration signal recognition is higher.
What problem does this paper attempt to address?