Crack the Code: Advanced Method Enhances Train Rivet Inspection

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.May 17 2024

In a recent study published in the journal Scientific Reports, researchers proposed a novel method for detecting cracks in train rivets using fluorescent magnetic particle detection (FMPFD) and instance segmentation. Their technique aimed to achieve high accuracy and recall in identifying cracks that are dense, multi-scale, and uninstantiated and may affect the performance and safety of railway vehicles.

Background

Train rivets are pivotal components of railway vehicles' connections and are prone to damage. Cracks in rivets can compromise their load-bearing capacity, elevating the risk of failure. Therefore, non-destructive testing methods are indispensable for assessing rivet quality and upholding railway transportation safety.

FMPFD is a widely used technique for detecting defects in metal parts. It works by applying a magnetic field to the part and spraying fluorescent magnetic particles on its surface. The particles are attracted by the leakage flux lines caused by the defects and emit a visible glow under ultraviolet light. However, manual inspection of fluorescent images is labor-intensive and susceptible to errors.

About the Research

In this article, the authors introduced an instance segmentation method for automated crack detection of train rivets using FMPFD. Instance segmentation involves detecting and segmenting each object in an image, irrespective of its category. They employed an enhanced version of you only look once version 5 (YOLOv5), a state-of-the-art deep learning-based object detection algorithm, for instance, segmentation of fluorescent cracks.

To address the specific challenges of the fluorescent crack detection task, the study made several improvements to the YOLOv5 algorithm, including:

• Implementing a decentralized target center and low overlap rate labeling method to reduce overlap between adjacent bounding boxes and assign cracks to different grid cells for better prediction and regression.
• Introducing a Gaussian-weighted correction post-processing method to reduce confidence in overlapping bounding boxes rather than eliminating them, preserving diversity and improving the recall rate.
• Incorporating an efficient channel spatial attention mechanism (ECSAM) to enhance channel and spatial attention of feature maps, capturing more discriminative features of the cracks.
• Utilizing a multi-task feature learning method that leverages both detection and segmentation datasets to learn semantic and spatial features of the fluorescent cracks, improving precision and reducing background interference.

Additionally, the paper described the development of a single coil non-contact train rivet magnetization device capable of magnetizing rivets with varying shapes and sizes and detecting both radial and axial cracks simultaneously. This device is user-friendly, portable, and offers advantages over conventional industry equipment.

The researchers evaluated the performance of their proposed method on a dataset comprising 480 train rivet fluorescent crack images, encompassing various crack distribution scenarios. They compared their method with mainstream object detection algorithms such as faster region-based convolutional neural networks (Faster-RCNNs), YOLOX, YOLOv6, YOLOv7, and YOLOv8, as well as different attention mechanisms including channel attention (CA), squeeze-and-excitation (SE), efficient channel attention (ECA), simple attention module (SimAM), convolutional block attention module (CBAM), and bi-directional feature pyramid network (BiFPN).

Research Findings

The outcomes showed that the proposed method outperformed other approaches in terms of precision, recall, and mean average precision (mAP), particularly in detecting dense, multi-scale, and irregular cracks. Specifically, the new method achieved a recall rate and mAP of 86.4% and 90.3%, respectively, significantly surpassing the baseline model. Furthermore, it attained a higher mean intersection over union (mIoU) compared to semantic segmentation algorithms such as a convolutional neural network with a U-shaped encoder-decoder structure (UNet).

The presented approach achieved a balance between speed and accuracy, achieving a frame per second (FPS) of 17.8 while maintaining a certain absolute speed. Additionally, the authors provided detection examples to illustrate the efficacy of their model in accurately locating, classifying, and segmenting each crack in the image, thereby reducing false positives and false negatives.

Applications

The proposed method has significant implications for train rivet crack detection and other industrial product inspection tasks. It has the potential to decrease labor intensity and the rate of missed detections associated with manual inspection, thereby contributing to improved product quality and safety.

Furthermore, its adaptability allows for generalization to various defect types and components, including welds, bearings, and crankshafts. Additionally, it can be applied to other non-destructive testing methods, such as X-ray and infrared thermography.

Conclusion

In summary, the novel approach demonstrated effectiveness in automated crack detection of train rivets, achieving high accuracy and recall even for dense, multi-scale, and uninstantiated cracks. It holds significant value for quality inspection in various industries beyond railway transportation, extending to metal parts and potentially other fields requiring crack detection. Moving forward, the researchers suggested the following directions for future work:

• Enhancing model robustness and generalization through expanded data collection and advanced data augmentation techniques.
• Fine-tuning post-processing or evaluation metrics to improve the regression accuracy of prediction boxes and enhance detection result interpretation.

Exploring alternative deep learning-based methods like generative adversarial networks (GANs) and graph neural networks (GNNs) while comparing their performance and advantages against the proposed method.