Utilizing an Instance Segmentation Network Capable of Balancing Precision and Speed to Achieve Fine-Grained Vibration Displacement Measurement of Rotating Bodies

Visual sensor vibration measurement technology demonstrates significant potential in the field of rotating body condition monitoring. To address the issues of object detection's inability to stably obtain the bounding box of rotating bodies over the long term and the semantic segmentation methods' inability to distinguish multiple targets of the same category, the article proposes a refined measurement method that balances accuracy and speed. The method applies instance segmentation networks to vibration measurement, effectively resolving the confusion in distinguishing multiple targets of the same category. Furthermore, it integrates the backbone of the YOLO series network with MLPBlock through residual nesting to ensure detection speed while accurately extracting the features of rotating bodies. A feature pyramid network with dynamic computation weights is then constructed to achieve the fusion of rotating body information, thereby improving segmentation accuracy. Additionally, Concat channels and coordinate attention (CA) modules are introduced to enhance the saliency of rotating body features and improve localization accuracy. A high-speed industrial camera is used to build a vibration dataset to measure the vibration displacement of single and multiple targets. By comparing with existing algorithms, this article verifies the superior performance of the proposed method in vibration displacement measurement. Notably, in terms of the key evaluation metric normalized root mean square error (NRMSE), the proposed algorithm achieves outstanding results of 0.2203 and 0.1744 in the X and Y directions, respectively, for single-target vibration displacement measurement. Moreover, the displacement curves obtained by this method exhibit the highest fitting degree with the eddy current signal curves. In multitarget measurement scenarios, the algorithm achieves NRMSEs of 0.2807 and 0.2722 for the left and right rotors, respectively, effectively distinguishing multiple rotating bodies of the same category and demonstrating its effectiveness and applicability in multitarget measurement scenarios. This study not only effectively addresses the problems encountered in object detection and semantic segmentation algorithms but also improves the accuracy of vibration displacement measurement of rotating bodies.