Automated identification of fissure trace in mining roadway via deep learning

Abstract

The fractured rocks distributed near the structural surface of coal roadway heading face can easily cause water or air conduits, ultimately resulting in accidents that threaten production safety, such as mine flooding, gas leakage, or supporting failure. Therefore, successful detection of fissures is crucial to mine safety. Despite the rising popularity of computer vision in accurate fissures detection, it fails to satisfy the demand of engineering practice. To address this problem, this paper first establishes a 1000-image database of fissures of coal roadway heading face based on data collection, cleaning, and annotation. Then, a framework for fissure detection and segmentation is constructed with deep convolutional neural network (DCNN) named DeepLabv3+ serving as the overall architecture, and a lightweight MobileNetV2, instead of the original Xception, as the main feature extraction network. The database is then employed to train and test the neural network model. Finally, the robustness and adaptability of the model under the common jamming environment in coal mines are evaluated. According to the results, the deep learning algorithm, which performs favorably in identifying various fissures in the coal roadway heading face, is immune to interference such as low illumination, wire mesh, multi-scale edge, cutting mark, and concentrated light beams. Notably, the performance of such method is on par with, or better than, humans in performing individual image segmentation. The approach performs higher segmentation accuracy and calculation speed than the traditional image identification algorithm, thus realizing rapid identification of fissures in coal mines in batch. This study can provide a reference for image semantic segmentation of fissure traces under similar conditions.

Keywords

• Intelligent driving
• Computer vision
• Fissure evaluation
• DeepLabv3+
• MobileNetV2