基于深度神经网络的电厂跑冒滴漏智能识别方法研究

Research on Intelligent Identification Method of Power Plant Leakage Based on Deep Neural Network

电厂设备复杂,容易发生跑冒滴漏问题,人工巡检存在发现滞后、人为疏忽、不能实时传达异常情况等问题。基于深度学习卷积神经网络、迁移学习和小样本学习技术,设计电厂异常状态智能识别报警系统,利用深度学习模型检测监控系统捕获的现场图片,识别常见的设备跑冒滴漏现象,准确并且及时地发出警告,以此提高电厂的安全监管和对意外事故的应急能力。采用相对成熟的YOLOv5作为目标检测网络基础框架,针对跑冒滴漏数据稀少问题,对网络结构进行优化并采用迁移学习与小样本学习方法来提高网络识别精度。结果表明,基于深度学习卷积神经网络的电厂异常状态智能识别报警系统,能够保持电厂异常状态识别的准确性和实时性。该系统可以实现自主全天候智能检测,及时推送报警信息,减少利用人力关注监控设备排查异常状态可能发生的疏漏,降低电厂运行维护成本,提高电厂的安全监管与对意外事故的应急能力。

The complex equipment of power plant is prone to the problem of leakage.There are some problems in manual inspection,such as delayed discovery,human negligence,and unable to convey the abnormal situation in real time.Based on the deep learning con⁃volutional neural network,transfer learning and few⁃shot learning technology,an intelligent recognition and alarm system for abnormal state of power plants is designed,deep learning model is used to detect the field pictures captured by the monitoring system,common equipment leakage is identified and warnings are given accurately and timely,so as to improve the safety supervision of the power system and the ability to respond to accidents.The relatively mature YOLOv5 is adopted as the basic framework of target detection network.Ai⁃ming at the problem of sparse leakage data,the network structure is optimized and the transfer learning and few⁃shot learning methods are adopted to improve the accuracy of network recognition.The results show that the power plant abnormal state intelligent recognition and alarm system based on deep learning convolutional neural network can keep the accuracy and real⁃time recognition of power system abnormal state.The system can realize autonomous all⁃weather intelligent detection,timely push alarm information,reduce the possible omissions because of the use of human attention monitoring equipment to check abnormal status,reduce the operation and maintenance costs of the power system,and improve the power system safety supervision and emergency ability.