| Most of the coal mines in China are underground mines.Restricted by geological and mining conditions,most coal mines have complex production systems,poor underground production environment,many operators,prominent hidden dangers,and easy accidents.In order to achieve safe production,safety supervision departments,technical departments and professionals have been set up to carry out safety supervision,inspection,and hidden danger investigation and management.A lot of human,financial and material resources have been invested,but the problem of safe production has not been completely solved.With the rapid development and wide application of modern information and communication technology,coal mine informatization work is gradually advancing from the digital mine and perception mine construction stage to the intelligent mine construction stage,which will not only provide strong technical support for coal mine safety production,but also will promote the innovation of coal mine safety management methods and patterns.At present,technologies such as automation,Internet of Things,Internet,cloud computing,and artificial intelligence have begun to be applied in coal mine production and management,resulting in a large amount of data and information such as numerical values,audio,video,and pictures.How to use these information resources to solve safety management problems such as intelligent identification and comprehensive management of hidden dangers in coal mines has become a focus of attention and research.Therefore,this thesis builds a training data set based on various data generated by various systems of intelligent mines through monitoring,perception and operation,designs two types of hidden danger identification models based on AI intelligent algorithms,and integrate them with cloud computing,big data and AI technology to build an integrated system that can be applied to the actual production environment.The details are as followsThis thesis analyzes the characteristics of big data of coal mine safety hazards,and divides it into video image data and time series data accordingly.On this basis,the characteristics of hidden dangers contained in different types of data are summarized,and the hidden dangers are classified.The hidden dangers in video images are divided into static hidden dangers,dynamic hidden dangers and complex type hidden dangers,and the hidden dangers in time series are divided into numerical prediction hidden dangers and classification hidden dangers.Then,preliminary identification method for these five types of hidden dangers is given.On this basis,the training datasets are constructed using various methods.The video image training data mainly includes public data on the Internet,data collected in the field,and artificial data generated by UE4 3D engines.This thesis uses Mix Up,Cut Mix,Mosaic and other methods to enhance the data sets.Time series training data mainly includes underground field data collected by various sensors,monitoring and perception systems in coal mines.The real time series data is simulated and synthesized using the Time Gan model.Through the enhancement and synthesis processing of video image training data and time series training data,the expansion of training data is realized.The intelligent identification model of different types of hidden dangers based on video image data are constructed.The dataset is augmented with data augmentation and data synthesis methods.On this basis,the Yolo X neural network is used to construct a static type hidden danger identification model,which realizes intelligent identification of hidden dangers such as no helmet,foreign body on the belt conveyor,fire,etc.The detection accuracy rate reaches 89.9%,and the Deep Sort algorithm is used to realize the hidden danger target tracking.Based on the Yolo X model,using the Alpha Pose attitude detection model and the ST-GCN action recognition model,a dynamic type hidden danger identification model is constructed,which realizes the intelligent identification of hidden dangers such as fighting,falling,and sleeping in the well,and the detection accuracy reaches 90.0%.The rule reasoning method is used to synthesize the above models and the Mono Flex monocular 3D target detection model to construct a hidden danger detection model for complex scenes,which realizes the detection of complex hidden dangers such as crossing the belt,illegally picking up vehicles,pedestrians and vehicles rush to the road and foreign body on belt conveyor within non-dedicated cameras.The detection accuracy of the above four hidden dangers reaches 88%,92%,90% and 89% respectively.Through testing,the data enhancement method used in this thesis improves the accuracy of Yolo X and Alphapose model by 2.8% and 2.2% respectively,and the synthetic data generated by the UE4 3D engine improves the accuracy of the four models of Yolo X,Alpha Pose,ST-GCN and Mono Flex by 4.3%,4.8%,6.6% and 3.2% respectively.The experimental results show that the data enhancement and data synthesis methods used in this thesis can effectively improve the training effect of the model.The intelligent identification model of different types of hidden dangers based on time series data are constructed.Using Time Gan-based data synthesis method to expand the data set and pre-training strategy to improve the model training effect.Taking gas concentration prediction and shearer overheating trip fault prediction as examples,the hidden danger identification model based on numerical prediction and the hidden danger identification model based on classification type are constructed by using LSTM,GRU and GPT models in NLP,and the corresponding hidden dangers are identified by using two kind of time series.The stacking method is used to fuse the three models,which further improves the accuracy of hidden danger identification.In the task of gas concentration prediction,the RMSE of the three models reaches 0.143,0.141 and 0.138 respectively,the MAPE reaches 1.12%,1.16% and 1.11% respectively,and the RMSE and MAPE of the fusion model reaches 0.131 and 0.98% respectively.In the shearer overheating trip fault prediction task,the accuracy rates of the three models reaches 84.47%,83.62%,and 85.4%,respectively,the F1-Score reaches 0.496,0.480,and 0.517,respectively,and the accuracy and the F1-Score of the fusion model has reached 89.15 and 0.595 respectively.After testing,the pre-training strategy and the synthetic data generated based on Timegan model improve the accuracy of stacking model in the prediction task of shearer overheating trip fault by 6.9% and 3.6%respectively,and F1 score by 0.156 and 0.074 respectively.The experimental results show that the pre-training and data synthesis methods used in this thesis can improve the training effect of the model.The constructed intelligent identification model of different types of hidden dangers based on time series data can use time series data to quickly and accurately identify hidden dangers in coal mines.The fusion model has better detection results than a single model.An integrated system is designed that integrates the intelligent identification model of hidden dangers based on video image data and time series data.In order to realize the practical application of the above constructed models,a set of intelligent identification system for coal mine safety hazards is designed and initially developed by integrating these models with cloud computing,AI and big data technologies.The bottom layer of the system is based on Kubernetes+Docker,which can flexibly configure and deploy each module of the system.Using the Flink big data framework,a data warehouse of hidden dangers in coal mines is constructed,which realizes the real-time aggregation and storage of data,and provides data source for the intelligent identification models.The overall identification process of hidden dangers is designed,and based on Nvidia Trinton Inference Server and Deep Stream technology,the online deployment and inference of the model are realized.The model is optimized based on Tensor RT technology,and the design of two sets of application layer APPs on the desktop and mobile terminals is realized based on the Vue framework. |
PDF Full Text Request