School of Mechanical Electric & Information Engineering,China University of Mining & Technology(Beijing),Beijing　 100083,China

Coal is the main energy source in China,and the coal mining industry is a high-risk industry. Coal mine flooding is one of the mine disasters in coal mines. The existing coal mine flooding monitoring and alarming methods have some problems,such as poor adaptability,high false alarm and missed alarm rate,which are difficult to meet the needs of coal mine safe production. It is an effective measure to detect floods in time and to block and drain water rea- sonably. Mine water level can be monitored by water level scale image. When the depth residual neural network is used to recognize the image,the recognition effect of the image is closely related to the depth of the depth learning network.In this paper,based on different depth recognition algorithms,the recognition performance of mine water level scale is analyzed and studied. The image of water level scale of working face and roadway is collected, and the image is marked,and the image database is established. The scale image scale center position parameters,shape size parame- ters,scale classification are extracted as feature vectors,and trained by residual neural network. When the network training is stable,the feature vectors are obtained by the same operation of the image to be detected,and the feature vectors are parsed into the key information of the image target to realize the scale target detection of the water level scale. Experiments are carried out for different network depths,and the loss rate and stability,average recognition rate, f1 value,PR curve,ROC curve,training time and testing time are compared. When the number of training pictures is fixed,the algorithm has the best depth,and too deep network will lead to inadequate training,and too shallow network will lead to over-fitting. The influence of confidence threshold on the average recognition rate is analyzed. When confi- dence is 0. 4,the average recognition rate is the highest. The recognition rate and time-consuming comparison of other common algorithms are also made. The average training time of this algorithm is 625 ms,and the average testing time is 47 ms. The recognition rate of mine water level scale calibration target is more than 97% .

国家重点研发计划资助项目(2016YFC0801800)

CAO Yuchao,FAN Weiqiang. Performance analysis and research of mine water level scale recognition based on differ- ent depth recognition algorithms[J]. Journal of China Coal Society,2019,44(11):3529-3538.