Spatially embedded transformer: A point cloud deep learning model for aero-engine coaxiality prediction based on virtual measurement

Abstract

Coaxiality is a critical indicator of assembly accuracy in aero-engines, directly impacting the device's operational performance and lifespan. Due to the enclosed nature of the aero-engine casing system, measuring the coaxiality of assembled components presents significant challenges. This paper introduces a novel deep learning architecture, the spatially embedded transformer (SETrans), designed to predict coaxiality from unassembled part data by correlating it with the contact surface points of assembled components. Additionally, a virtual measurement model is developed to collect micron-scale point cloud data, facilitating the fine-tuning of the deep learning model. The SETrans utilizes the transformer's capability for global information aggregation to process point cloud inputs, capturing the comprehensive relationships across assembled surfaces. A newly designed module, the spatial bias, integrates distance and angular information between neighboring point clouds into the transformer block, enhancing the model's ability to capture fine-grained local details. Experimental validation is 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 the SETrans across diverse measurement techniques. Comparative analysis with other point cloud deep learning benchmarks shows that the proposed approach achieves top prediction accuracies of 93.65% and 94.31% with a coaxiality precision of 0.01 mm across different data domains. These results confirm the effectiveness of the SETrans and demonstrate its adaptability to real-world assembly conditions involving various components.