💻 computer science

Coal Mine Drill Pipe Detection Algorithm Integrating Large-Kernel Separation and Coordinate Locking Mechanisms

This paper proposes a lightweight YOLO11n-based drill pipe detection algorithm for complex underground coal mine environments that integrates Large Separable Kernel Attention and Coordinate Attention to enhance feature robustness against occlusion and lighting variations, alongside an InterpIoU loss to improve regression accuracy for slender targets, achieving superior generalization, reduced miss rates, and high real-time inference speeds.

Original authors: Canjun Xiao, Zaisheng Jiang, Chenyu Wang, Junxiang Yang, Siliang Zheng, Youping Li, Dongdong Chen, Shengrong Xie

Published 2026-10-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Canjun Xiao, Zaisheng Jiang, Chenyu Wang, Junxiang Yang, Siliang Zheng, Youping Li, Dongdong Chen, Shengrong Xie

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the earth's surface, in the dark and dusty tunnels of a coal mine, a critical operation is underway. Workers and machines drill deep into the rock to release trapped gas, a necessary step to prevent explosions and make mining safer. To know if the job is done correctly, they must count exactly how many long, thin metal pipes are pushed into the ground. If the count is wrong, the gas might not be drained properly, leaving a deadly hazard. For decades, this counting has relied on human eyes or sensors that break easily in the harsh environment. Now, researchers are turning to cameras and artificial intelligence to solve this problem, hoping to give the mines a reliable, unblinking eye that can see through the chaos of the underground world.

The challenge, however, is that these underground tunnels are not like a clean laboratory. They are filled with swirling dust, sudden bursts of blinding light from welding or lamps, and heavy machinery that constantly blocks the view. The pipes themselves are long and slender, often looking like thin lines that can easily disappear behind a pipe rack or a worker's arm. Standard computer vision systems, which are usually trained on clear, well-lit photos of everyday objects, struggle immensely here. They tend to get confused by the noise, losing track of the pipes or mistaking a shadow for a pipe. When these systems fail, the automated counting stops, and the mine must rely on manual checks, slowing down the work and reintroducing human error.

A team of researchers from universities and coal mining companies in China has developed a new way to help computers see these pipes clearly. They started with a modern, lightweight detection system—a type of software designed to find objects in images quickly. But they realized that the standard version was not built for the specific headaches of a coal mine. To fix this, they redesigned the software's "brain" to pay attention to the right things. Instead of just looking at small patches of an image, they taught the system to look at the whole picture at once, connecting broken pieces of a pipe even when a machine or a cloud of dust cuts it in half. They also added a special mechanism that locks onto the physical edges of the pipe, helping the system ignore the glare of a bright light or the confusion of a dark corner.

The researchers tested their new system in a very rigorous way. Rather than just showing it pictures from the same mine where it was trained, they tested it on completely different mines with different lighting, different equipment, and different layouts. This approach, known as testing on "out-of-distribution" data, is like teaching a student to drive in a quiet neighborhood and then immediately testing them on a busy, rainy highway to see if they can really handle the road. The results showed that their improved system was far more reliable than the standard versions. It successfully found the pipes in difficult conditions where other systems failed, reducing the number of missed pipes by a noticeable margin.

In terms of raw numbers, the new system detected the pipes with an accuracy of 88.7 percent in these tough, unseen environments, a significant jump from the baseline. It also became better at counting the pipes correctly in video footage, improving the final count accuracy by over nine percentage points in the most challenging video tests. Perhaps most importantly for a real-world mine, the system remained fast and efficient. It did not require massive, expensive computers to run; it could process over 1,300 images every second, which is far faster than the speed of a standard video camera. This means the system can run on the small, rugged computers that are already installed in the mines, providing real-time feedback without slowing down the drilling process.

The researchers also discovered that their changes worked best when the pipes were partially hidden or when the light was extreme. In tests involving blinding glare or pipes blocked by other machinery, the new system kept its focus on the metal rod, while older systems would often lose the target entirely. However, the team noted that the system still faces limits when the pipes are extremely far away and appear as tiny specks in the image, a problem that will need more work to solve. Despite this, the study proves that by tailoring artificial intelligence to the specific, messy reality of the underground, we can create tools that are not just smart, but truly robust enough to work where humans cannot always be. This advancement brings the coal industry one step closer to fully automated, safer drilling operations, where the count is always right, and the danger is always lower.

Technical Summary: Coal Mine Drill Pipe Detection Algorithm Integrating Large-Kernel Separation and Coordinate Locking Mechanisms

Problem Statement
Underground coal mine drilling operations require accurate, continuous counting of drill pipes to evaluate gas drainage compliance and borehole pressure relief. However, traditional contact-based sensing fails under harsh conditions, while existing machine vision solutions struggle with the specific environmental challenges of underground mines. These environments are characterized by significant dynamic variations, frequent target occlusions by pipelines and machinery, intense light glare, localized low illumination, and pervasive dust. Conventional object detection models, particularly those trained on standard datasets, suffer from severe overfitting and rapid generalization degradation when deployed in these unknown, dynamic working faces. Specifically, the slender geometry of drill pipes makes their boundary features fragile; they are prone to physical truncation and photometric distortion, leading to high miss rates and unstable bounding box regression in slender targets.

Methodology
The authors propose a lightweight drill pipe detection algorithm based on the YOLO11n network, specifically engineered to address the generalization and regression challenges of slender targets in complex environments. The approach focuses on two core architectural pathways: feature enhancement and bounding box regression.

Feature Enhancement (LSCA Module):
At the end of the backbone network, the standard C2PSA module is replaced with a Large Separable Kernel Coordinate Attention (LSCA) module. This module integrates two parallel branches to decouple background noise and enhance topological connectivity:
- Branch A (Large-Kernel Spatial Attention): Utilizes Large Separable Kernel Attention (LSKA) with decomposed $1\times k$ and $k\times 1$ convolutions (with dilation) to expand the receptive field. This bridges physical occlusion boundaries, allowing the network to connect fractured features of drill pipes obscured by pipelines or machinery.
- Branch B (Coordinate Attention): Employs Coordinate Attention (CA) to generate height and width masks. This branch performs orthogonal pooling to precisely lock onto the physical edges of the drill pipes, resisting interference from intense light and dust.
- Fusion: The outputs of both branches are fused via a single element-wise multiplication with the original feature map, preserving high-frequency edge details while injecting dual priors for cross-occlusion connectivity and edge locking.
Flexible Regression Loss (InterpIoU):
To address the sensitivity of bounding box regression to extreme aspect ratios (common in slender drill pipes), the authors introduce an Interpolation-based Flexible Regression Loss (InterpIoU). Standard losses often cause gradient oscillations when the short edge of a predicted box deviates slightly, leading to box jitter. InterpIoU constructs a transitional interpolated box ( $B_i$ ) between the predicted box ( $B_p$ ) and the ground truth ( $B_g$ ) using a coefficient $\lambda=0.5$ . The loss function replaces the standard overlap term with the Intersection over Union (IoU) of this interpolated box, while retaining center-distance and aspect-ratio penalties from CIoU. This creates a smooth gradient transition that suppresses short-edge divergence and stabilizes localization.
Dataset and Validation Strategy:
The study utilizes a dataset collected from nine distinct drilling sites. Crucially, the authors employ an Out-of-Distribution (OOD) validation strategy based on spatiotemporal physical isolation. The test set is constructed from sites and timeframes strictly separated from the training data to objectively evaluate the model's true generalization capability on unseen working faces, rather than relying on random data splits.

Key Contributions

LSCA Mechanism: The integration of large-kernel separation and coordinate locking into a parallel multiplicative fusion module specifically tailored for the occlusion and glare characteristics of underground drill pipe scenes.
InterpIoU Loss: A novel regression loss tailored for slender rigid bodies that mitigates gradient oscillations caused by extreme aspect ratios, improving bounding box compactness.
OOD Evaluation Framework: The construction of a rigorous validation set based on spatial and temporal isolation to reflect real-world deployment scenarios where models face unknown dynamic conditions.
Lightweight Architecture: The proposal maintains a lightweight profile (2.561 M parameters) suitable for edge deployment while significantly outperforming baseline models in accuracy.

Experimental Results
Experiments were conducted on an NVIDIA RTX 5880 Ada GPU using a dataset of over 6,000 images (expanded to ~14,500 via non-destructive augmentation). The improved model was compared against mainstream lightweight detectors (YOLOv5n, v8n, 11n, 12n, 26n).

Generalization Performance: On the OOD test set, the proposed model achieved 88.7% mAP@50 and 43.7% mAP@50-95, representing improvements of 2.0% and 2.2% over the baseline YOLO11n, respectively.
Anti-Interference Capability: Under complex conditions (glare, occlusion, dust), the miss rate was reduced by 1.4% (to 18.3%), and recall increased by 1.4% (to 81.7%). Layered analysis showed significant gains in bright glare conditions (precision increased by 5.4 percentage points) and medium highlight scenarios (miss rate reduced by 3.8 percentage points).
Efficiency: The model maintains a high inference speed of 1303.8 fps with 2.561 M parameters and 6.4 GFLOPs. While slightly slower than the baseline (1556.1 fps), it remains well above the frame rates of standard underground surveillance cameras.
Counting Accuracy: When coupled with the ByteTrack tracker, the improved model increased counting accuracy on five challenging video cases by 9.1 percentage points compared to the baseline (average 78.7% vs. 69.7%), demonstrating robust support for automated counting logic.

Significance
The paper claims that the proposed algorithm provides a robust, readily deployable visual perception solution for automated drill pipe operations in underground coal mines. By addressing the specific failure modes of slender target detection in dynamic, high-interference environments, the method overcomes the generalization limitations of conventional models. The integration of the LSCA module and InterpIoU loss effectively suppresses background noise and stabilizes localization, laying a foundation for the intelligent and unmanned development of mining equipment. The authors note that while the method significantly improves performance, distant small targets remain a current boundary, suggesting future work involving higher-resolution inputs or dedicated shallow detection heads.

More like this