2. Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China;
3. College of Control Science and Engineering, Zhejiang University, Hangzhou, Zhejiang 310058, China
Overview: As a new type of sensing technology, the phase-sensitive optical time domain reflection (Φ-OTDR) distributed optical fiber sensing technology has the advantages of good environmental tolerance, low energy consumption, high sensitivity, long monitoring distance compared with traditional sensing technology, and has been the emphasis and hotspot of the researches. Among the manifold application fields of Φ-OTDR distributed optical vibration sensing system, identification of vibration events is one of the most popular and advantaged applications. Therefore, efficient and accurate identification of different vibration events is the focus of this paper. The recognition speed of traditional methods is fast, but their recognition rate and robustness are not ideal and strongly dependent on the artificial feature design. Although the method based on two-dimensional convolutional neural network (2-D CNN) has obtained a high recognition rate, the slow recognition speed and feature extraction process hamper its use in real-time systems. In view of the fact that one-dimensional convolutional neural network (1-D CNN) has been used in other real-time identification fields and achieved good results, this paper proposes a multi-scale one-dimensional convolutional neural network (MS 1-D CNN) method which takes the vibration signals as the input of this network and needs not to manually extract features. Feature extraction of the intrusion vibration signals in the MS 1-D CNN takes into account the rich feature information of the signals in time and frequency scales, thus achieves efficient and accurate identification. In order to control the spatial complexity and parameter quantity, three scales and four layers are used in the MS 1-D CNN method. The raw vibration signals are pre-processed firstly to remove noise as far as possible, including pre-emphasis filtering, normalization and spectral subtraction. The pre-processed waveform signals are directly inputted into the MS 1-D CNN, and the recognition results are achieved by using fully-connected layer (FC layer) and Softmax layer. In comparison with the methods based on 2-D CNN and 1-D CNN, the proposed method balances the time and frequency scales well during feature extraction and reduces the number of pending parameters. A vibration recognition experiment was designed to classify the three different vibration events, including damaging, knocking and interference. The recognition results show that MS 1-D CNN achieves similar recognition performance at twice processing speed compared to 2-D CNN. Hence, it is beneficial to improve the real-timing of vibration monitoring while maintaining the recognition performance.