Mining Subsidence Disaster Warning Research Based On CNN

In the mining process,affected by various natural or human factors,mining subsidence occurs from time to time,and in serious cases,even threaten people’s lives and property safety,so it is of great significance to carry out early warning of mining subsidence.Traditionally,most of the mining subsidence observation methods use geodetic techniques such as total station measurement,precise leveling rod and GPS,but these methods are susceptible to climatic reasons and have problems such as expensive and difficult operation and construction.In recent years,deep learning technology has been effective in the field of image processing,and the application of convolutional neural network and image processing to mining subsidence disaster early warning can provide powerful support for coal power mining disaster early warning,so this paper proposes a research of mining subsidence disaster early warning method based on convolutional neural network.This paper firstly pre-processes the mining test site extraction images.By collecting 3000 mining subsidence images at the mining test site,the HSV color feature extraction,combined with morphological filtering,edge detection and other methods to de-noise and anti-interference the collected feature images,so as to realize the pre-processing of mining images;then the images are divided into five different subsidence disaster warning levels according to the size of the subsidence cracks appearing on the mining surface at different moments.The data augmentation method is used to further expand the mining subsidence disaster early warning experimental dataset,which has a total of 13,920 images after augmentation.A deep learning convolutional neural network approach is introduced,i.e.,the MobileNetV1,MobileNetV2,and MobileNetV3 networks are used for mining subsidence disaster warning experiments.Subsequently,the MobileNetV3 model was improved and a MobileNetV3_SAM network model was proposed,and the network models of MobileNetVl-V3 and MobileNetV3_SAM network models were compared for early warning analysis of mining subsidence images,and the results showed that the accuracy of each network model for early warning prediction of subsidence hazards reached 85%above,and the improved MobileNetV3_SAM network has the highest accuracy of 93%for sinkhole disaster prediction.On the other hand,to solve the problem of high time overhead of training network models in deep learning,a training strategy combining migratory learning and Cosine Warmup is proposed,namely Transfer learning_CW.The trained VGG network models,MobileNet net models,and EfficientNet models from the datasets in deep learning experiments and ImageNet datasets are selected and EfficientNet models,retrain the trained networks using Transfer learning_CW,and conduct experimental comparison of mining subsidence disaster warning,the results show that,compared with the traditional training methods,using the transfer learning method can improve the training speed by 5-14 times and the model accuracy by 1-1.2 times,while using the Transfer learning_CW method,it can further improve the training speed by about 20%-30%and the model accuracy by about 1%-3%further.Meanwhile,the prediction accuracy of each model under this training mode reaches over 91%,with the EfficientNet model predicting up to 98%.Finally,in order to observe the sinkhole status of the mining face more intuitively,the location information of the sinkhole monitoring points in the mining face is fed back graphically in real time.Figure[64]Table[17]Reference[36]...