基于多尺度双阶段网络航空发动机涡轮叶片故障检测研究

摘要:

工业相机在图像采集过程中受到各种噪声的干扰,如电子噪声、热噪声等。这些噪声会在图像中表现为随机分布的亮点或暗点,影响图像的清晰度和质量。进而导致故障检测的FPR-MRCI值下降。对此,研究基于多尺度双阶段网络航空发动机涡轮叶片故障检测方法。首先,机器视觉利用工业相机与镜头、环型LED光源及异形夹持装置,通过中空旋转平台带动叶片旋转实现全方位图像采集。然后,针对采集到的图像利用多尺度双阶段网络进行超分辨率重建。双阶段设计将特征提取和上采样分开处理,在特征提取阶段深入挖掘图像的本质特征,上采样阶段则专注于图像的重建和细节优化,从而有效提高图像的分辨率和质量,解决图像噪声问题。最后,通过本体建模构建涡轮叶片故障检测知识图谱,实现系统性的故障检测知识整合。构建基于特征嵌入的涡轮叶片故障检测卷积神经网络模型,将实体向量与超分辨率重建图像输入特征嵌入模块,实现涡轮叶片故障检测。测试结果表明。设计系统对于检测难度较高的晶界腐蚀、颗粒磨损以及气动磨损故障的检测结果均正确。设计系统的FPR-MRCI整体高于0.6,说明设计系统具有较高的故障特征识别率,同时误报率相对较低。随着噪声强度的增加,设计系统的下降幅度较小,说明其涡轮叶片故障识别性能在受到噪声影响后仍然比较准确。

Abstract:

Industrial cameras are subject to various types of noise interference during image acquisition, such as electronic noise, thermal noise, etc. These noises will appear as randomly distributed bright or dark spots in the image, affecting the clarity and quality of the image. This leads to a decrease in the FPR-MRCI value for fault detection. Regarding this, research is conducted on a multi-scale two-stage network based fault detection method for aircraft engine turbine blades. Firstly, machine vision utilizes industrial cameras and lenses, circular LED light sources, and irregular clamping devices to achieve omnidirectional image acquisition by driving blade rotation through a hollow rotating platform. Then, a multi-scale two-stage network is used for super-resolution reconstruction of the collected images. The two-stage design separates feature extraction and upsampling processing. In the feature extraction stage, the essential features of the image are deeply explored, while in the upsampling stage, the focus is on image reconstruction and detail optimization, effectively improving the resolution and quality of the image and solving the problem of image noise. Finally, a knowledge graph of turbine blade fault detection is constructed through ontology modeling to achieve systematic integration of fault detection knowledge. Constructing a convolutional neural network model for turbine blade fault detection based on feature embedding, inputting entity vectors and super-resolution reconstructed images into the feature embedding module to achieve turbine blade fault detection. The test results indicate that. The design system provides accurate detection results for grain boundary corrosion, particle wear, and aerodynamic wear faults that are difficult to detect. The overall FPR-MRCI of the design system is higher than 0.6, indicating that the design system has a high fault feature recognition rate and a relatively low false alarm rate. As the noise intensity increases, the decrease in the design system is relatively small, indicating that its turbine blade fault identification performance is still relatively accurate even after being affected by noise.