Crack Identification and Flaw Detection Eva-luation of Bolster Hanger Based on Machine Vision
［Objective］ The bolster hanger, as a critical component of the bolster spring suspension system, is highly susceptible to fatigue cracks due to excessive dynamic loads and prolonged service life, significantly impacting railway operational safety. To meet the intelligent magnetic particle flaw detec...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Urban Mass Transit Magazine Press
2025-02-01
|
| Series: | Chengshi guidao jiaotong yanjiu |
| Subjects: | |
| Online Access: | https://umt1998.tongji.edu.cn/journal/paper/doi/10.16037/j.1007-869x.2025.02.027.html |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | ［Objective］ The bolster hanger, as a critical component of the bolster spring suspension system, is highly susceptible to fatigue cracks due to excessive dynamic loads and prolonged service life, significantly impacting railway operational safety. To meet the intelligent magnetic particle flaw detection requirements for bolster hangers and assist inspection personnel in the flaw detection work, special research on crack identification and flaw detection evaluation of bolster hangers is conducted based on machine vision. ［Method］ To address the limitations of the YOLOv5 algorithm backbone network in capturing crack information, it is proposed to integrate the SimAM (similarity-aware attention module) mechanism into the backbone network to enhance the model′s sensitivity to crack information and its resistance to background noise interference synchronously. In addition, to overcome the potential information loss during the feature fusion stage of the Neck network, an enhanced BiFPN (bidirectional feature pyramid network) structure is introduced for efficient fusion of multi-scale feature maps. By implementing a weighted fusion strategy and bidirectional connection mechanism, the loss of critical bottom-level information is effectively reduced. ［Result & Conclusion］ The performance of the YOLOv5-SA-BF model is tested using an image acquisition system for bolster hangers. Experimental results demonstrate that the improved algorithm achieved a MAP (mean average precision) of 98.08%, representing a 2.91% increase over the original model, and a Recall increase of 4.13%. The model effectively addresses background false positives and low detection accuracy issues, meeting the requirements in actual inspection. |
|---|---|
| ISSN: | 1007-869X |