Gas Time Series Prediction And Anomaly Detection Based On Deep Learning

With the increase of the intensity and depth of coal mining in China,gas outburst accidents have occurred occasionally which has seriously threatened and restricted the safety production of coal mines and the sustainable development of the coal industry and also endangered the national economy and people's life safety.The gas outburst accident does not happen suddenly.Some signs will be released before the occurrence of the gas outburst,so the warning of gas outburst accidents can be realized according to the released signs.The change of gas concentration is one of the many precursors.The gas concentration and its changes contain a lot of prominent information,which embodies the three factors of gas outburst: ground stress,gas pressure,coal structure and it is also the easiest to monitor.Under the support of the National Natural Science Foundation of China,combined with the specific coal mine project,this paper obtains data from the coal mine monitoring system,and uses deep learning methods to solve gas concentration prediction,gas concentration state identification and gas outburst factors estimatation.And then proposes a gas outburst accident warning model.The main achievements are summarized as follows:1.A Long Short-Term Memory Network model applied to the prediction of gas concentration time series is constructed.Long Short-Term Memory Network,a special kind of special Recurrent Reural Network,is a type of deep learning model.This model has the advantage of being able to capture long-term dependency information between time series data,and overcomes the gradient disappearance and gradient explosion problems of RNN,which makes it very suitable for time series prediction.Therefore,according to the characteristics of gas time series,LSTM is applied to gas concentration time series prediction.This model can not only make short-term prediction but also make long-term prediction,which provides a basis for gas concentration status recognition and gas outburst warning.2.A LSTM based on Grey Wolf Optimizer is proposed.The global convergence of GWO is used to optimize the parameters of Long Short-Term Memory Networks,which overcomes the problem of convergence to the local optimal solution of BPTT used in LSTM.The model is applied to the prediction of gas time series.Compared with BPTT-based LSTM,the new model does not require a large amount of data,and can obtain better performance under a small amount of data,and the convergence speed is faster than the original model.3.A multi-dimensional time series classification model based on Cycle_DBN is established.This model inherites the advanced representation learning ability of DBN and introduces a feedback loop in the DBN to capture the dependence between time series data.The identification of the gas concentration state cannot only look at the gas concentration value,but also needs to consider other parameters.The gas concentration,wind speed and temperature are selected to identify the gas concentration state.The time series classification model based on Cycle_DBN is used to identify the gas concentration state.Simulation experiments show that the model is more suitable for time series classification than Deep Belief Network model.4.An adaptive gas concentration state recognition model is designed.The model can continuously collect real-time information of coal mines during the operation process,and dynamically adjust the gas concentration state identification model parameters,so that the model can work in a better state.The simulation results show that the performance of the adaptive gas concentration state recognition model is improved compared with the original model.5.An early warning model for gas outburst is prosposed.Since the change of gas concentration is one of the key signs before gas outburst,it reflects the three elements of gas outburst,namely ground pressure,gas content and coal structure.Therefore,the three elements of gas outburst can be calculated according to the change of gas concentration.Finally,the comprehensive index of gas outburst is concluded from the three indicators.According to the comprehensive index value,the state of the working face can be evaluated,and the gas outburst warning model is constructed accordingly.