Structural displacement monitoring based on mask regions with convolutional neural network

Abstract Displacement monitoring is an important part of structural health monitoring. This paper proposes a structural displacement monitoring method based on mask regions with convolutional neural network (Mask R-CNN), which combines deep learning with machine vision to extract the coordinates of calibration object. First, a calibration object dataset was collected. The dataset contains 165 images. Second, the dataset was trained by Mask R-CNN to obtain a monitoring model, which can generate a mask to cover the calibration object in the image. Third, the mask information is extracted from the image to obtain the coordinates of calibration object. Finally, the structure displacement can be obtained by the coordinates. In this paper, the method is applied to the static displacement monitoring and dynamic displacement monitoring. The results showed that average error of static displacement monitoring is 2.05%, and the correlation coefficient between monitoring results of the proposed method and laser displacement sensor monitoring results is as high as 0.9836 in dynamic displacement monitoring. To solve the problem of remote displacement monitoring, a telephoto lens was mounted on a smartphone to monitor displacement of 10 m distance. The results showed that the average error of remote displacement monitoring is 4.18%.