近年来，煤炭行业的安全生产形势尤为严峻，煤矿事故导致的死亡人数居高不下。其中，人为因素是导致煤矿事故频发的首要原因。随着智慧矿山概念的兴起与发展，矿工不安全行为的识别研究取得了一定进展。然而，煤矿环境的复杂性和特殊性使得当前的研究仍面临三大挑战。首先，矿工不安全行为的数据集获取难度极大。目前公开的数据集中，尚未有专门针对矿工不安全行为的数据集。其次，矿工不安全行为识别网络的准确性有待提高。在地下矿井中，光线条件往往不佳，加之粉尘、水雾等因素的干扰，使得矿井环境变得异常复杂，这无疑对识别网络的准确性造成了直接影响。最后，矿工不安全行为识别网络的实时性也是一个亟待解决的问题。由于现存的深度学习模型的复杂性，现有识别网络往往难以做到实时响应。针对以上问题，本文在现有理论的基础上，对矿工不安全行为识别的研究内容如下：

首先，针对矿工不安全行为缺乏公开大型数据集的问题，本文采用了矿用防爆摄像仪在矿井仿真实验室的不同场景中进行了视频采集。这些场景包括了采煤工作面、掘进工作面、运输巷道，3个具有代表性的煤矿作业环境。在采集过程中，特别关注了摔倒、不戴安全帽、违规扒乘、跨越设备以及闯入危险警戒区，共5类行为，并辅助数据增强技术扩充数据集，完成煤矿井下不安全行为数据集的构建。

其次，针对现有检测方法对矿工不安全行为识别准确率较低的问题，提出了一种改进YOLOv8的矿工不安全行为识别算法。并将SimAM注意力机制融入到颈部网络的C2f模块后，由能量函数计算的通道权重中，推导出针对每个神经元的三维注意力权重，以增强对井下环境特征的表达能力。在与不同注意力机制对比实验中mAP@0.5达到93.69%、精度达到94.62%、召回率达到94.52%、FLOPs达到8.9G。

然后，针对现有检测方法对矿工不安全行为识别实时性较低的问题，将YOLOv8作为基础框架，引入轻量级EfficientViT骨干网络，该结构采用级联分组注意力模块与“夹层布局”策略，在保证检测精度的同时降低计算复杂性，提高算法的实时性。在不同骨干网络的对比实验中参数量为12.4M、mAP@0.5达到94.93%、推理时间为13.4ms、FLOPs为1.5G。同时改进损失函数为EIoU，将回归聚焦于在高质量的先验框，以加速模型收敛，增强网络的实时性。在不同损失函数对比实验中mAP达到94.73%，精度为93.89%，召回率为94.42%。该模型兼顾了实时性和准确性，能够满足实际项目应用需求。

最后，为满足实际煤矿工作场景下的需求，结合改进YOLOv8的行为识别方法，基于PyQt5实现实验平台的设计，利用自制数据集对系统的性能验证，该平台在煤矿作业场景下表现优秀，具备了准确识别和实时响应的特点，为煤矿智能化建设和人员安全保障提供了有力的技术支撑。

In recent years, the safety situation in the coal industry has been particularly severe, and the number of deaths caused by coal mine accidents remains high. Among them, the human factor is the primary cause of the frequent occurrence of coal mine accidents. With the rise and development of the smart mine concept, the research on the identification of miners' unsafe behaviours has made some progress. However, the complexity and specificity of the coal mine environment make the current research still face three major challenges. First, it is extremely difficult to obtain datasets of miners' unsafe behaviours. Currently, there is no dataset specifically targeting miners' unsafe behaviours in the publicly available datasets. Second, the accuracy of the miners' unsafe behaviour recognition network needs to be improved. In underground mines, light conditions are often poor, and the interference of dust, water mist and other factors make the mine environment extremely complex, which undoubtedly has a direct impact on the accuracy of the recognition network. Finally, the real-time nature of the miners' unsafe behaviour recognition network is also an urgent problem. Due to the complexity of existing deep learning models, it is often difficult for existing recognition networks to achieve real-time response. Aiming at the above problems, this paper, based on existing theories, researches on miners' unsafe behaviour recognition as follows:

Firstly, to address the lack of publicly available large-scale datasets on miners' unsafe behaviours, this paper used a mining explosion-proof camera to capture video in different scenarios in the mine simulation laboratory. These scenarios include coal mining face, digging face, and transport roadway, three representative coal mine operating environments. During the acquisition process, special attention was paid to falling, not wearing a helmet, illegal pickpocketing, crossing equipment, and intruding into the dangerous warning area, a total of five types of behaviour, and assisted data enhancement technology to expand the dataset, completing the construction of the dataset of unsafe behaviours in underground coal mines.

Secondly, to address the problem of low accuracy of existing detection methods in identifying miners' unsafe behaviours, an algorithm to improve YOLOv8 for miners' unsafe behaviours identification is proposed. And after incorporating the SimAM attention mechanism into the C2f module of the neck network, the three-dimensional attention weights for each neuron are derived from the channel weights calculated by the energy function to enhance the representation of the underground environment features. In comparison with different attention mechanisms, mAP@0.5 reaches 93.69%, the accuracy reaches 94.62%, and the recall rate reaches 94.52% and FLOPs reach 8.9G.

Then, to address the problem of low real-time recognition of miners' unsafe behaviours by existing detection methods, YOLOv8 is used as the basic framework, and a lightweight EfficientViT backbone network is introduced, which adopts the cascading grouped-attention modules and the "mezzanine layout" strategy to reduce the computational complexity and improve the real-time performance of the algorithm while guaranteeing the detection accuracy. The structure adopts cascading grouped attention modules and "sandwich layout" strategy, which ensures the detection accuracy while reducing the computational complexity and improving the real-time algorithm. In the comparison experiments of different backbone networks, the number of parameters is 12.4M, the mAP reaches 94.93%, and the inference time is 13.4 ms and FLOPs is 1.5G. Meanwhile, the loss function is improved to be EIoU, and the regression is focused on the high-quality prior frames to accelerate the convergence of the model and enhance the real-time performance of the network. The mAP reaches 94.73%, the precision is 93.89%, and the recall is 94.42% in the comparison experiments with different loss functions. The model balances real-time and accuracy, and can meet the requirements of real project applications.

Finally, in order to meet the needs under the actual coal mine working scenarios, combined with the improvement of YOLOv8's behaviour recognition method, the design of the experimental platform is implemented based on PyQt5, and the performance of the system is verified by using the homemade dataset, which performs well under the coal mine operating scenarios with the features of accurate recognition and real-time response, and provides a powerful technical support for the intelligent construction of coal mines and the safety and security of the personnel.