煤岩截割状态识别是实现采煤工作面“无人化”开采的关键技术，为了实现煤岩截割状态信息的实时感知与精准判别，结合虚拟样机技术，提出基于CPS（Cyber Physical Systems）理念的煤岩截割状态识别方案，将煤岩截割状态信息的获取、处理、识别等异构数据进行多领域融合。开发不同赋存条件的煤岩离散元模型，建立采煤机截割部刚柔耦合虚拟样机模型，利用DEM-MFBD（Discrete Element Method-Multi Flexible Body Dynamics）双向耦合技术确保运动信息与煤岩状态特征信号数据的实时传递，获取采煤机截割煤岩的振动信号，并通过STFT（Short-Time Fourier Transform）算法将其转化为二维时频图像。结合时频域信息特征，实现煤岩截割状态信息识别模块的搭建。构建一种基于DCGAN-RFCNN（Deep Convolutional Generative Adversarial Networks-Random Forest Convolutional Neural Networks）网络模型的煤岩截割状态识别方法，通过使用改进的DCGAN网络进行时频图像的扩充，采用增加梯度惩罚项的方式提升合成样本维持原始样本特性的能力，生成每类仿真工况包含5 000个合成样本的煤岩时频图像数据集，将仿真原始数据集与合成样本数据集混合作为煤岩截割状态识别网络的训练集与测试集，采用改进的RFCNN算法对模型进行训练，得到模型识别结果。选取不同数量合成样本的数据集以及不同识别方法的网络模型进行对比分析，结果表明，当RFCNN识别网络中未添加合成样本时，其平均识别率为89.74%，随着合成样本数量的增加，煤岩截割状态的识别率提升，当添加合成样本数量达到5 000时，识别效果最佳，平均识别率达到98.09%，验证了采用改进的DCGAN网络丰富数据集的优越性。RFCNN网络模型与CNN，PSO-BP，BP网络模型相比收敛速度快、泛化能力强、识别率高，在煤岩截割状态识别中效果显著，可对软岩硬煤、夹矸层较多等复杂赋存条件做出准确判断。通过构建的煤岩时频谱图像数据集对DCGAN-RFCNN网络进行试验验证，利用混淆矩阵统计煤岩截割状态的识别率为98.41%，与仿真结果接近，验证了该方法的可行性。依托Simulink仿真平台，成功构建基于CPS感知分析的煤岩截割状态识别系统，实现实时数据共享、在线感知与控制，使采煤机具备智能截割能力。

The recognition of cutting state of coal-rock is the key technology to realize “unmanned” mining in coal face. In order to realize a real-time perception and accurate judgment of coal-rock cutting state information, combined with virtual prototype technology, a coal-rock cutting state recognition scheme based on CPS (Cyber Physical Systems) was proposed. It integrated heterogeneous data such as coal-rock cutting state information acquisition, processing, recognition and so on in multiple fields. The discrete element models of coal-rock with different occurrence conditions were developed. The rigid flexible coupling virtual prototype model of shearer cutting part was established. Using the DEM-MFBD (Discrete Element Method-Multi Flexible Body Dynamics) two-way coupling technology to ensure the real-time transmission of motion information and coal and rock state characteristic signal data, the vibration signal of shearer cutting coal and rock was obtained and converted into some two-dimensional time-frequency images by STFT (Short-Time Fourier Transform) algorithm. Combined with the characteristics of time-frequency information, the module of coal and rock cutting state information recognition was built. A method of coal-rock cutting state recognition based on the DCGAN-RFCNN (Deep Convolutional Generative Adversarial Networks-Random Forest Convolutional Neural Networks) network model was constructed. By using improved DCGAN network to expand the time-frequency image, and the gradient penalty term was added to enhance the ability of composite samples to maintain the characteristics of the original samples. The coal-rock time-frequency image data set with 5000 composite samples in each simulation condition was generated. The original simulation data set and composite sample data set were mixed as the training set and test set of coal-rock cutting state recognition network. The improved RFCNN algorithm was used to train the model and get the recognition results. The data sets of different numbers of synthetic samples and the network models of different recognition methods were selected for contrast analysis. The results show when no composite sample is added to the RFCNN identification network, the average recognition rate reaches 89.74%. With the increase of the number of composite samples, the recognition rate of coal-rock cutting state is improved. When the number of composite samples is 5000, the recognition effect is the best, and the average recognition rate reaches 98.09%, which verifies the superiority of using improved DCGAN network to enrich data sets. Compared with the CNN, PSO-BP, BP network models, the RFCNN network model has fast convergence speed, strong generalization ability and high recognition rate, and has significant effect in coal-rock cutting state recognition. It can determine complex occurrence conditions such as soft rock hard coal and more gangue layers. The DCGAN-RFCNN network was experimentally verified through the constructed coal and rock time-frequency spectrum image data set, and the recognition rate of the state of coal rock cutting calculated by using the confusion matrix was 98.41%. The results are close to the simulation results, which verify the feasibility of the method. With the Simulink simulation platform, the coal and rock cutting state recognition system based on CPS perception analysis has been constructed to realize real-time data sharing, online perception and control, so that the shearer has the ability of intelligent cutting.

1 基于CPS感知分析的煤岩截割状态识别系统框架

   1.1 物理系统搭建

   1.2 信息系统构建

2 煤岩截割状态信息获取

   2.1 煤岩离散元模型建立

   2.2 截割部刚柔耦合虚拟样机模型建立

   2.3 截割部DEM-MFBD双向耦合模型建立

3 煤岩截割振动信号分析与处理

   3.1 振动信号特征提取

   3.2 数据处理

4 基于DCGAN-RFCNN网络的煤岩截割状态识别

   4.1 改进DCGAN模型的样本扩充

   4.2 RFCNN模型训练

   4.3 模型的对比验证与分析

   4.4 试验验证

5 基于CPS感知分析的煤岩截割状态信息反馈系统的构建

6 结论