Abstract
The assembly quality of precision instruments is closely linked to their operational quality and lifespan. Inadequate assembly can lead to instruments failing to meet required accuracy standards post-assembly, necessitating disassembly and reassembly. Substandard assembly may also degrade the performance of the equipment and, in some cases, lead to fatigue fractures, causing severe safety incidents. To ensure the working quality and lifespan of precision instruments while reducing assembly costs, this paper utilizes deep learning to predict and evaluate assembly accuracy during three stages: pre-assembly, post-assembly, and during operation. Prior to assembly, deep learning is employed to predict assembly errors to avoid the need for reassembly due to such errors. After assembly, assembly errors are analyzed to ensure instrument quality. During operation, the assembly status is monitored to ensure equipment longevity. This paper systematically explores the precision assembly of precision instrument critical components such as engine casings and rotors, providing new theories, methods, and technologies for ensuring the reliability of precision instruments. The main research efforts are as follows:(1) A deep learning model based on point cloud inputs has been proposed to predict the assembly accuracy of aerospace engines. By analyzing the point cloud data from the contact surfaces of engine casings, this model can directly predict the coaxiality post-assembly, thereby preventing the occurrence of significant errors that lead to disqualification and the need for reassembly. During the research, a coordinate measuring machine was utilized to capture precise point cloud data of the engine casings. Subsequently, these data were upsampled using Non-Uniform Rational B-Splines to construct a deep learning dataset that correlates assembly point clouds with coaxiality. This dataset was then fed into a novel point cloud deep learning backbone network—Self-Channel Cross-Attention Point Network. Through this network, end-to-end prediction of coaxiality based on the point clouds of aerospace engine surfaces is achieved, significantly enhancing the efficiency and accuracy of assembly precision prediction.
(2) A new point cloud deep learning backbone, SETrans, has been developed to handle different sampling methods and data distributions in assembly scenarios. In this research, SETrans leverages the global information aggregation capabilities of transformers to process point cloud inputs, capturing the comprehensive relationships between assembly surfaces. A newly designed module, the spatial bias, integrates distance and angular information between neighboring point clouds into the transformer block, thereby enhancing the model's ability to capture fine-grained local details. Experimental validation was conducted using two distinct datasets representing different assembly scenarios: the aero-engine casing, sampled using contact-based coordinate measuring machines, and the rotor, sampled using non-contact optical gaging products. These specific sampling methods test the generalizability of SETrans across diverse measurement techniques.
(3) A monocular vision measurement model based on keypoint detection has been proposed to evaluate the assembly quality of fully assembled precision instruments. This method facilitates the measurement of the exposed length of bolt connections within the assembly, which is used to assess the quality of the assembly and ensure its proper functioning. Initially, an image acquisition platform is constructed to collect images of bolts. These images are processed to extract the region of interest (ROI), and the camera parameters are calibrated. Subsequently, the ROI is input into the deep learning model, which identifies seven keypoints using a heatmap. The camera calibration parameters are then transferred to the monocular vision measurement model, establishing the relationship between different coordinates. Using the 2D keypoints, 3D models of the bolt’s upper and lower surfaces are constructed. Finally, the spatial distance between these two surfaces is determined, enabling the calculation of the bolt's exposed length for quantitative detection of the assembly state.
In summary, this work utilizes point cloud deep learning to construct a mapping relationship between the contact surfaces of precision instrument assemblies and their assembly errors. This approach enables direct end-to-end prediction of assembly coaxiality based on the distribution of point clouds on contact surfaces. Additionally, through visual measurement and keypoint detection, this research facilitates monitoring of the operational status of precision instruments after assembly, significantly aiding in the assurance of their functioning and the estimation of their lifespan. The algorithms proposed in this thesis have been validated through physical experiments and have undergone robustness tests with various datasets, ensuring the generalizability of the proposed system.
| Date of Award | 7 Apr 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Steven WANG (Supervisor) & Jun Liu (External Co-Supervisor) |