The arrival of big data era puts forward new requirements for the application of power video surveillance． The power video monitoring system nowadays only plays the role of remote video recorder basically． The capability of video monitoring system in active early warning is still not used，and the substation is still not " intelligent enough" ． Based on the convolutional neural network technology，the original video material is extracted from the existing video monitoring platform． The video cloud storage platform is used for storage and video2pic is used for data cleaning． Then people's unsafe behavior of the scene is classified，analyzed and marked． Finally，the intelligent upgrade of the existing video monitoring is realized through feature extraction，classifier model training and verification． Through the mining of large video data，real － time intelligent early warning analysis of the situation on the spot is realized，and automatic mining，real － time early warning and intelligent push of irregular operation and business hall service behavior of substation are realized． By getting rid of the dependence on manual duty，saving labor investment and improving power production safety control capacity and quality service level，the power video monitoring system can give full play to its " supervision" and " control" role．