<p>The intelligent detection of computer chassis assembly states is crucial for ensuring quality control and improving production efficiency in large-scale computer manufacturing processes. To overcome the limitations of traditional methods in dealing with complex internal backgrounds, fine-grained assembly features, and frequent component occlusions, this paper proposes a lightweight detection framework, YOLO-FGA (You Only Look Once-Fine-Grained Anomaly), and presents an optimized design tailored to industrial computer vision applications. The model integrates the Re-parameterized Gradient Efficient Layer Aggregation Network (RepGELAN) in the YOLOv11 backbone structure, significantly enhancing the ability to extract features for subtle assembly differences. Additionally, a novel Context-Aware Attentive Screening Feature Pyramid Network (CASA-FPN) is introduced in the Neck structure, which resolves feature misalignment caused by occlusions and complex backgrounds via adaptive multi-scale fusion. Furthermore, a channel-wise knowledge distillation (CWKD) strategy is employed to enhance detection robustness while maintaining computational efficiency. Evaluation on a dataset containing 15 chassis components demonstrates that the YOLO-FGA model, after incorporating knowledge distillation, achieves significant performance improvements compared to YOLOv11n: a 1.1% increase in mAP₅₀, a 1.2% increase in mAP₅₀_:95, a 4.1% increase in Acc, a 1.4% increase in F1-score, and a 30% reduction in number of parameters. These results demonstrate the potential and effectiveness of the method in high-precision and resource-efficient quality inspection systems for assembly states.</p>