Study On The Intelligent Monitoring System Of Pump Room In Coal Mine Based On Machine Vision

Operational status of underground drainage system affects the stability and safety of coal mine production.As a key part of drainage system,pump house is the high-level restricted area in the underground mining system.To ensure safe operation in pump house,it is necessary to obtain the abnormal working state of the pump in time,and to restrict the workers' access to danger area.Nowadays,the available monitoring systems provide simple support services including on-site working conditions monitoring and video storage,but these systems cannot meet the need of early-warning and intelligent monitoring.To address these problems,several techniques such as machine vision and deep learning are employed to achieve intelligent online monitoring of pump house.This research mainly includes four aspects as follows:To achieve automatic water leakage detection,a strategy,the combination of moving target detection and the convolution networks,is proposed.First,coterminous frames differencing is applied to determine the target area which is considered as the potential leaking area.Afterwards,convolution network is trained to further investigate the characteristic of the potential leaking area.Experimental results show that the Resnet-50 model yields the best performance with an accuracy of 96.8%.The Deepsort model is employed to achieve multi-target tracking.First,three algorithms including YOLOv3,YOLOv3-tiny,and YOLOv3-MobileNet are trained as the detector of Deepsort model.To extract feature information,ReID is employed and trained according to the loss in cosine similarity.Finally,the detector and ReID network are combined to build a Deepsort model to track of the workers in the pump house.The experimental results show that the Deepsort model based on YOLOv3 detector achieves the highest multiple objects tracking accuracy of 83.7%.To recognize the actions and behaviors of people in pump house,3D-CNN is employed to identify six basic actions.First,the dataset is comprised of the cropped images with a fixed window.Second,three models including C3 D,ResNet-3D and Inception-3D model are introduced to recognize the actions.In addition,an efficient PI3 D model,which combines P3 D and Inception-3D networks,is proposed.The number parameters of PI3 D network are reduced by 58%,compared with the number parameters of standard Inception-3D model.Furthermore,PI3 D achieves an accuracy of 91.5%.Based on the above algorithms,an intelligent online monitoring software is designed to provide well-rounded support services including water leakage warning,staff trajectory tracking and behavior recognition.