基于迁移学习与残差网络的矿工不安全行为识别

Identification of miners’ unsafe behaviors based on transfer learning and residual network

为精确识别矿工的不安全行为,降低煤矿事故发生率,提出迁移学习结合深度残差网络的图像识别方法。将矿工的行为样本划分为完全安全行为、较安全行为、不安全行为3个维度,其中完全安全行为包括:走路、坐下、站立,较安全行为包括:弯腰、下蹲、抬东西、推、拉、挥手、拍手,不安全行为包括:跌倒、投掷;采用ResNet50网络进行训练,微调ImageNet数据集迁移学习的权重参数,通过全连接层进行12分类,并将最终分类结果与测试数据进行对照检验。研究结果表明:基于迁移学习的残差网络模型识别跌倒与投掷动作的准确率,优于其他深度神经网络模型,能够有效识别不安全行为从而避免由人为因素导致的事故发生。

In order to accurately identify unsafe behaviors of miners and reduce occurrence of accidents in coal mines,an image recognition method combining transfer learning and deep residual network is proposed.Firstly,behavior instances of miners were divided into three dimensions,namely completely safe behaviors,relatively safe behaviors,and unsafe behaviors,among which completely safe behaviors included walking,sitting and standing,relatively safe behaviors included bending,squatting,lifting,pushing,pulling,waving and clapping,and unsafe behaviors included falling and throwing. Then,Res Net50 network was used for training,and transfer learning weight parameters of Image Net data set were fine-tuned. Finally,12 classification was conducted through full connection layer,and final classification results were checked against test data. The results show that residual network model based on transfer learning is superior to other deep neural network models in identification accuracy of falling and throwing movements,and it can effectively identify unsafe behaviors,thus avoiding accidents caused by human factors.