Fracture Localization of Structure Based on Transfer Learning and Broadband MUSIC Algorithm

To enhance the resilient protection of underground engineering, real-time monitoring of cracks in the whole engineering is of great significance. In the present paper, a real-time localization method for structural cracking is proposed. Through the method of transfer learning, the picking model of structural cracking direct acoustic wave and reflected acoustic wave was established after training the deep learning model of seismic phase-pick with 1847 sets of training data. Automatic recognition and extraction of cracking acoustic signals is realized in this way. To solve the problem that the localization accuracy of broadband multi-signal classification (MUSIC) algorithm is reduced due to spatial aliasing, the equation between the direction of sound source and inter channel phase difference is established through the relationship of time delay in spatial and frequency domain. The narrow bands with two specified positive phase differences are used to calculate spatial spectral in localization after further analysis of the relationship between phase difference and spatial aliasing. The in-situ lining cracking simulation test verifies that the improved algorithm effectively reduces the spatial aliasing and improves the positioning accuracy. The average error distance between the estimated cracking point and the actual cracking point is 0.05 m. The monitoring of the extending direction and length of crack can be realized after connecting each cracking point, which can meet the engineering needs.