Using the YOLOv12 Model for Verifying the Correct Color Sequence of Wires in Network Cables (Patch Cords) on the Production Line

Technical Summary: Using the YOLOv12 Model for Verifying the Correct Color Sequence of Wires in Network Cables

Problem Statement
In the industrial production of network cables (patch cords), the correct arrangement of eight wire strands within the RJ45 connector is critical for network performance. Standards such as T568A and T568B define precise color sequences; any deviation results in defective, non-functional products. Traditional quality control relies on manual visual inspection using digital microscopes. The authors identify this method as time-consuming, tedious, and prone to human error due to operator fatigue, variable lighting conditions, and limitations in visual acuity. These limitations necessitate an automated solution that can operate in real-time on high-volume production lines to ensure uniform quality and reduce waste.

Methodology
The study proposes an intelligent, automated inspection system based on the twelfth version of the YOLO (You Only Look Once) object detection model, specifically YOLOv12. The methodology encompasses the following components:

• Dataset: A dedicated dataset of 2,500 high-resolution images (approx. 1200×1600 pixels) was captured using an industrial digital microscope from the front view of RJ45 connectors. The dataset includes eight distinct color classes corresponding to the T568B standard (white-orange, orange, white-green, blue, white-blue, green, white-brown, and brown).
• Data Preparation: Images were manually annotated with bounding boxes for each wire. The data was split into 70% training, 15% validation, and 15% testing. To improve generalizability and prevent overfitting, data augmentation techniques were applied, including rotation, cropping, zooming, horizontal flipping, and adjustments to brightness and contrast.
• Model Architecture: The system utilizes YOLOv12, a single-stage detector enhanced with attention mechanisms. Key architectural features employed include:
  • Region-based attention in the backbone to focus on relevant image areas.
  • Residual-ELAN structures for feature aggregation to preserve information at network depth.
  • Multi-scale prediction to detect thin wires at three different output levels.
  • Anchor boxes: Nine anchor boxes were optimized via k-means clustering based on actual wire dimensions.
• Training Configuration: The model was trained using the PyTorch framework on an NVIDIA Tesla T4 GPU. The loss function was a weighted combination of localization (CIoU), objectness, and classification losses (weighted 0.5, 0.4, and 0.1, respectively). Training utilized a batch size of 16 over 50 epochs with early stopping mechanisms.
• Decision Logic: Post-detection, a sorting algorithm arranges the detected wires based on their horizontal (x-axis) coordinates. This predicted sequence is then compared against the standard T568B pattern to determine if the cable passes or fails.

Key Results
The system was evaluated on a completely independent test set of 375 images under varied conditions (lighting, noise, background). The performance metrics are as follows:

• Precision: 99.1%
• Recall: 98.0%
• Mean Average Precision (mAP@IoU 0.5:0.95): 94.6%
• Sequence Verification Accuracy: 98.13% (368 out of 375 cables correctly classified as pass/fail).
• Error Analysis: The model exhibited high accuracy for solid-color wires. The few errors (less than 2%) occurred primarily with wires featuring white insulation and narrow colored stripes (e.g., white-orange, white-brown) due to color similarity and low contrast. Notably, the system produced zero "false accept" errors (defective cables marked as good), with all 7 errors being "false rejects" (good cables marked as defective).
• Speed: The system processes images at approximately 160 frames per second (5–6 ms per image) on a GPU, and roughly 8–9 fps on a CPU, significantly exceeding the speed requirements of industrial production lines.

Significance and Claims
The paper claims that the proposed YOLOv12-based system offers a superior alternative to both traditional manual inspection and earlier automated methods (such as the 2022 Tsai & Cheng approach, which was limited by lighting dependency and slow processing speeds).

The authors assert that the system successfully addresses the trade-off between speed and accuracy, delivering real-time performance suitable for Industry 4.0 environments. By automating the verification process, the system eliminates human error, reduces production bottlenecks, and ensures 100% quality control. The study concludes that this approach is not only effective for network cables but is also extensible to other multi-strand inspection tasks, such as automotive wire harnesses and electronic board connections, representing a significant step toward fully automated smart factories.