In MR-assisted systems for industrial maintenance, accurately identifying equipment parts is fundamental for fault diagnosis and repair guidance. However, traditional part detection methods suffer from low accuracy as well as poor real-time performance, notably in complex backgrounds, under variable lighting conditions, and when detecting small targets with weak texture features. To tackle aforementioned issues, this paper proposes an optimized part detection algorithm for MR applications, YOLO-C2iE, based on an improved version of the YOLOv11 model. Firstly, ConvNeXt V2 is introduced into the backbone network, which can extract richer and more representative features at a lower computational cost, thereby reducing the consumption of computing resources and ensuring real-time processing speeds. Secondly, a lightweight attention mechanism, iEMA, is added. By integrating local perception with dynamic multi-scale modeling capabilities, it enhances the detection capability for small targets, reducing the occurrence of missed and false detections. The model was finally tested on a self-built dataset, experimental results show that compared to the original YOLOv11 model, YOLO-C2iE has achieved significant improvements across multiple key metrics, reaching 97.6% mAP (an increase of 4.2 percentage points over YOLOv11), with inference time reduced by approximately 47.9%, and FLOPs lowered to 5.9G. These results not only demonstrate the effectiveness of YOLO-C2iE in high-precision part detection but also showcase its outstanding real-time performance. This research provides a practical solution for part detection in industrial MR applications.