Research On Risk Prediction And Assessment Of Coal Mine Roof Based On Deep Convolutional Models

Coal mine roof hazards occur frequently,causing lots of losses.The reseach on prediction and assessment of roof risk commits an urgent task.The conventional methods cannot meet the actual needs of Chinese coal industry,while deep learning based models are able to make up their deficiencies.In the realm of coal mine safety,there is lack of deep learning based work.Especially in the topic of roof risk prediction and assessment,the relavent effort is still blank.Therefore,it is an urgent need for bringing deep learning into the field of coal mine safety,analyzing the coal mine monitoring data in nonlinear and chaotic ways.In particular,the improved models and algorithms which based on deep leaning are need to be carried out.Then,they need to be applied on predicting the local points and areas of underground coal mine roof.Furthermore,the proposed deep learning models need to used on evaluating the overall risk levels of coal mine roof dynamically.In this dissertation,we did sufficient literature reviews and methodology research.At the beginning,one chaotic analysis method has been studied,it foucs on the decoupling of multi-source dynamic monitoring coal mine data.Based on the foundation of chaotic analysis work,we proposed two roof risk prediction used generative models,which based on deep temporal convolutional neural networks.Those two deep learning models were applied to predict the local monitoring points of underground roof.Especially,they accomplished their tasks from the view of multivariate time series regression.From another idea,we abstraced the risk prediction of underground roadway roof local area as an imbalanced multivariate time series classification task.In this perspective,we proposed a deep convolutional classification model.What is more,considering the multi-source characteristic of coal mine roof risk assessment,we studied a deep convolutional feature fusion model to fuse the multi-source information and integrate the various monitoring variables for the dynamic evaluating the whole roadway roof area risk level.This paper maily contains the above contents:(1)Time series characteristics and chaotic analysis of coal mine roof monitoring variables.Under the chaos theory framework,we took nonlinear time series characteristics of the monitoring data into account and abstract the stress changing of coal mine roof as a chaotic system.We first validated the chaotic discrimination of coal mine roof system in a quantitative way.Based on the monitoring data,the multi-variable chaotic phase space reconstruction was applied to restore the dynamic characteristics of the original system.A maximum independent cross-correlation algorithm for multivariate phase space reconstruction parameters calculation was proposed.The proposed algorithm can effectively solve the high coupling problem for multi-monitoring variables.(2)Coal mine roof risk prediction based on deep temporal convolutional generative networks.In this part,deep learning has been brought into coal mine roof topic.To solve the long-term historical dependence problem in multivariable coal mine roof monitoring time series prediction,we studied the frameworks of deep recurrent network and deep convolutional network and proposed two generative models.Namely,deep dilated causal temporal convolutional network(DCTCNN)and long short-term memory recurrent network combined with convolution network(CNN-LSTM).These two generative models solved the long-term related historical problem from the ideas of expanding local receptive filed of CNN and combining CNN with RNN.The experimental results showed that DCTCNN and CNN-LSTM are superior to other algorithms,it can accomplish the local monitoring points risk of coal mine roof.(3)Coal mine roof hazard recognition based on imbalanced monitoring data.From the perspective of classification,we abstracted the problem of local area hazard recognition into an unbalanced classification of multivariable time series.An adaptive cost-sensitive learning algorithm was proposed,which can be used for training deep temporal convolutional neural networks and forcing them effectively classify imbalanced multivariable time series.Via the proposed algorithm,the temporal convolutional network(CNN),temporal fully convolutional network(FCN),temporal residual network(Res-Net)and temporal long short-term memory fully convolutional network(LSTM-FCN)were modified into cost-sensitive models.The experimental results illustrated that those models which were modified by the proposed algorithm are superior to other traditional algorithms.Finally,it is concluded that the adaptive cost-sensitive learning algorithm can effectively improve the performance of deep temporal convolutional networks and it is suitable for large-scale imbalanced time series classification task.The proposed method can effectively recognize the local area hazards of coal mine roof.(4)Coal mine roof risk assessment via deep convolutional feature fusion model.Based on the research of coal mine roof local points and local areas risk prediction,we also explored risk dynamic assessment of the whole roadway roof.A deep convolutional assessment model was proposed,it can fuse multi-source roof dynamic monitoring data and consider their spatial-temporal relationship.For simulating the temporal sequence proximity,periodicity and trend characteristics,we divided the monitoring data into different intervals with three time scales.In addition,the overall spatial position of roadway roof was used as training data labels and feed into the proposed model.We committed the experiment on a whole year historical monitoring data of a coal mine in 2017.The results demonstrated that the proposed deep convolutional assessment model can complete more detailed dynamic risk assessment work.What is more,it can achieve end-to-end learning from top board monitoring data to coal mine roof overall location and risk levels.