Reinforcement Learning And Optimal Tracking Control Of Multi-Time-Scale Interconnected Systems With Applications

Currently,most of the existing video monitoring systems in mines are aimed at monitoring the operating status of equipment and environmental conditions.Although there are many underground monitoring video data that play a certain preventive role in the safety production management of mining areas,they can only achieve the collection,transmission,storage,and display of video data,and lack the intelligent identification,analysis,and early warning functions for personnel and other moving target images,In addition,the complex underground environment(dim light,high camera position)leads to poor image quality,resulting in incomplete and untimely monitoring of moving targets on the underground working surface.Safety accidents often occur due to unsafe human behavior.Therefore,the identification and behavior analysis of people in mining areas and the automatic safety warning of irregular behaviors are the technical basis for realizing intelligent video safety monitoring in mining areas.In response to the above issues,this thesis proposes a visual recognition analysis method for peoples in mining areas.Starting from two aspects of mine personnel re-identification and mine personnel behavior analysis,it uses data augmentation,cross modal,and domain adaptive models to solve the problems of low visibility of underground imaging,poor generalization of training models,and uniform uniformity that are difficult to distinguish.On the basis of person re-identification and behavior analysis in complex mining environments is implemented,providing decision support for safe production in mining areas and promoting the development of underground work towards intelligence and standardization is a practical application of artificial intelligence in the field of mine industry,which has strong reference significance.The specific research content is as follows:(1)Aiming at the characteristics of uniform miner uniforms that are not easily distinguishable,this thesis discusses the attribute information applicable to miner data.Inspired by the person splicing and augmentation task,this thesis proposes a new data augmentation scheme based on segmentation and reorganization for miner data,and constructs reasonable supervision information for the augmented data.Then this chapter proposes a miner dataset with identity information and attributes annotations for a real mine environment,with a number of 34 miners and 171 images of miners.This work fills the gap in the academic miner dataset and lays a data foundation for subsequent domestic and foreign related research.Aiming at the high consistency of clothing among miners,an effective data augmentation strategy and label allocation scheme are proposed.Extensive experiments have shown the effectiveness of the methods proposed in this thesis.(2)Aiming at the problem of low imaging clarity of underground surveillance cameras in mining areas,this thesis proposes a visible infrared cross modal person re-identification method based on grayscale enhancement exploration and full modal center triplet.To address the significant modal differences and person appearance differences in visible-infrared re-identification,this thesis increases the diversity of training data,while seeking the optimal grayscale enhancement ratio to learn a more focused model.In addition,a full-mode central triad loss with stronger feature measurement capability is proposed to make the same person feature more compact and different person features distant from each other.Experiments on the common dataset demonstrate the effectiveness and superiority of the proposed method.(3)Aiming at the problem of poor generalization of person re-identification training models,in order to achieve arbitrary migration of trained personnel re-identification models to mining test sets,this thesis proposes a domain adaptive personnel re-identification method based on dynamic fusion optimization.Starting with the style differences between data sets,Mining area image enhancement and SPGAN mapping methods are used to convert source domain images into target domain image styles,and migrate person styles under different data sets,realize adaptive person re-identification in the field of dynamic fusion optimization,increase the number of training samples,improve the robustness of the recognition model to changes in the dataset,and improve the accuracy of personnel re-identification in mining areas.(4)In order to achieve personnel behavior recognition in mining areas,this thesis proposes a human posture estimation method for people in mining areas based on the hourglass attention high-resolution network.Firstly,we suppress the light effect on the image to obtain a mining area image with uniform lighting.Then,an hourglass attention feature extraction module is constructed based on deep high-resolution representation learning to supplement the feature aggregation process from high to low for high-resolution features.Then,a feature retrieval module is designed to remap the thermal map output from the task layer to a feature map of the same size as the output from this stage.The skeleton points learned from the previous stage of the network are returned to the next stage of the network for mutual supervised learning,Finally,a multi-stage supervision algorithm is designed to achieve the supervised training of high-resolution networks by combining relay supervision and self-distillation.Good human posture evaluation results have been achieved in both standard datasets and real scene applications in mining areas.The thesis contains 42 figures,14 tables,and 183 references.