基于深度学习的井下环境异常工况智能识别技术研究

Research on intelligent recognition technology of abnormal operating conditions in underground environment based on deep learning method

为了使装备在井下复杂环境中具有自主精准辨识井下对象和作业环境的能力,实现无轨装备及其他各类型装备无人化作业和开采,提出了基于利用深度学习方法开展井下环境异常工况智能识别分类研究。通过分析装备正常运行和作业的主要影响因素,构建了井下环境异常工况数据集,采用旋转变换、平移变换、缩放变换等数据增强技术,有效防止网络训练过拟合;基于InceptionResnet V2模型采用层冻结方法,重新训练全连接模型,通过不同的迁移策略进行实验对比分析。研究结果表明:添加2层全连接层,且每层包括4 096个神经元的迁移策略模型性能最佳,鲁棒性好,能够精准识别分类井下环境异常工况。

In order to make the equipments own the ability to identify the underground objects and operating environment independently and accurately in the underground complex environment, and realize the unmanned operating and mining of trackless equipments and other types of equipments, the research on intelligent recognition and classification of abnormal op- erating conditions in the underground environment based on the deep learning method was carried out. The dataset of abnor- mal operating conditions in the underground environment was established through analyzing the main factors affecting the nor- real running and operation of equipments, and the data enhancement technologies such as rotation transformation, translation transformation, zoom transformation and others were applied to effectively prevent the network training from overfitting. The fully connected model was retrained by using the layer freeze method based on the InceptionResnetV2 model, and the experi- mental comparative analysis was conducted with different transfer strategies. The results showed that the transfer strategy model with adding two-layer fully connected layer and containing 4096 neurons in each layer had the optimal performance with a good robustness, and it could identify the abnormal operating conditions in the underground environment accurately.