A visual measurement algorithm for vibration displacement of rotating body using semantic segmentation network

• Published in: Expert systems with applications Vol. 237; p. 121306
• Main Authors: Chai, Shanglei, Wang, Sen, Liu, Chang, Liu, Xiaoqin, Liu, Tao, Yang, Rongliang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2024
• Subjects: Deep learning; Rotating body; Semantic segmentation; Vibration displacement measurement; Deep learning; Rotating body; Semantic segmentation; Vibration displacement measurement
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2023.121306

Compared with the traditional vibration displacement measurement methods, visual vibration measurement offers several advantages such as long-distance capability, non-contact operation and easy installation. However, the phenomenon of low fitting accuracy of the bounding box often occurs when detecting rotating objects, resulting in a slight deviation in the relative offset of the center vibration point of the target between frames, which will cause a serious deviation in the regression of vibration displacement offset. In this paper, a high-speed industrial camera is employed as the image acquisition medium, and a deep learning-based semantic segmentation method is introduced to address visual vibration measurement challenges in rotating body. Specifically, the CSP module integrates different depth semantic information which is introduced into the Mobiledets backbone network in a targeted manner. This is not only strengthens the performance of the network for segmenting vibration objects, but also dramatically improves the practical performance of the algorithm. The conventional Relu activation function is substituted with Mish activation function, making the network more adept at segmenting rotating body in challenging backgrounds with varying illumination, blur, and similarity. The CSP+Mobiledets backbone network constructed in this study outperforms the U-Net network in terms of feature extraction effectiveness. Adding Dice-loss to the original loss function can more effectively solve the severe imbalance problem of samples caused by long-distance image acquisition. We take the most representative rotating body-rotor as the experimental subject. The displacement curve obtained by the existing algorithm has the best degree of fit with the signal curve collected by the eddy sensor. The results of different segmentation algorithms and detection algorithms on time domain curve plot, frequency domain plot and axis orbit plot are collectively compared. Furthermore, the results also provide valuable guidance for visual measurement of the vibration displacement of the rotating body in specific industrial scenarios. •A semantic segmentation network is applied for vibration displacement measurement.•CSP+Mobiledets backbone network to enhance feature extraction are proposed.•The proposed network has excellent performance under complex background.•The proposed algorithm has broad applicability on rotating structure.