Research On Tailings Pond Monitoring Method Based On Deep Learning

The construction of green mines helps to reduce negative impacts on the natural environment,reduce waste and excessive exploitation,extend the service life of mineral resources,and contribute to the sustainable utilization of mineral resources.Target detection of large-scale tailings ponds helps to protect the ecosystem of the area,timely detect and deal with potential environmental risks,maintain regional ecological balance,and reduce the degree of ecosystem damage.Identifying potential environmental issues,such as illegal dumping and discharge of tailings,can help detect anomalies in a timely manner before the problem escalates.Rapid response helps to reduce the likelihood o f environmental problems spreading and minimize damage to ecosystems.Intelligent monitoring of single tailings ponds can detect potential dangers in time,thereby improving the safety of tailings ponds,and is of great significance to ensuring the safety,environmental protection and efficient operation of tailings ponds.Deep learning methods have very powerful feature extraction and prediction performance,reduce the requirements for data feature description,and provide new solutions for processing mass ive data.Deep learning methods have made major breakthroughs and are used as advanced models in the field of artificial intelligence,which is a current research hotspot.In view of the current low intelligent level of tailings pond monitoring and the high reliability required for data-driven monitoring and prediction,this paper starts from the perspective of air-space-ground integrated monitoring of mine tailings ponds,and also considers the target detection of large-scale regional tailings ponds,dail y infiltration line monitoring and prediction of tailings ponds,and3 DWEB visualization.Supported by Sentinale-2 satellite images,tailings pond soakage line data monitored by ground real-time sensors,and drone aerial photography data.Taking Bayannur City in Inner Mongolia as a case study area,we explore the effectiveness of the deep learning method-two-stage convolutional neural network in large-scale tailings pond target detection.A typical tailings ponds-Huogeqi Western Copper Oubiliqi Tailings P ond as a case study area,the applicability of the two-way cyclic long and short memory network in monitoring and predicting the infiltration line of a single tailings pond was studied,and attempt to visualize 3DWEB for monitoring and predicting the infil tration line of the tailings pond.The aim is to explore an air-space-ground integrated module tailings pond monitoring method to provide an effective,intelligent monitoring and management service for the green development of future mining areas.The main research contents and conclusions are as follows:(1)Based on the two-stage convolutional neural network in deep learning methods,Sentinel-2 images were used to detect large-scale tailings pond targets in Bayannur District,Inner Mongolia.Firstly,coar se classification was performed on the tailings pond area and non-tailings pond area in the images,followed by fine-grained recognition and localization.Finally,fine detection was performed on tailings ponds in tailings area blocks with precise geograph ical coordinates.Research has shown that this method can automatically identify and accurately locate tailings ponds in remote sensing images,providing methodological support for intelligent detection and recognition of large-scale tailings pond targets and the management of illegal tailings discharge and storage.(2)Taking the typical tailings pond-Huogeqi Western Copper Oubiliqi Tailings Pond as a case study area,based on deep learning methods-bidirectional cyclic long short memory network,this s tudy investigates a tailings pond infiltration line prediction model with univariate input(infiltration line),bivariate input(infiltration line and internal displacement),and trivariate input(infiltration line,internal displacement and dam displaceme nt).Through model training and optimization,Comparison of sensitivity of different optimization methods to models,comparing multi-layer perceptron models with infiltration line regression prediction models based on bidirectional recurrent long short-term memory networks.Research has shown that in the prediction of infiltration line related problems in tailings ponds,the RMSE of multi-layer perceptron model,univariate input model,bivariate input model,and trivariate input model are 0.10611,0.09966,0.11955,and 0.11952,respectively.All four models have certain applicability for predicting a single tailings pond and win time in monitoring and prediction.(3)Based on drone low altitude photography and the Cesium engine,a 3DWEB visualization method for predicting and monitoring infiltration lines was explored using the Huogeqi Western Copper Oubiliqi Tailings Pond as an example.The results show that this meth od is simple,economical,and practical,and can solve the problem of weak two-dimensional visualization in monitoring the infiltration line of tailings ponds.The interface loading is smooth,and the presentation of the original single tailings pond is ma intained.It has certain applicability.