基于Faster R-CNN的旋转机械红外检测与识别

Infrared Detection and Identification of Rotating MachineryBased on Faster R-CNN

旋转机械是机械设备的核心部件,一旦发生故障会造成不可估量的损失,因此旋转机械的实时监测诊断显得尤为必要。无人值守的红外智能监测诊断将是故障诊断新的发展方向,要实现红外智能监测诊断首先要准确识别旋转机械部件。本文利用红外热像仪监测旋转机械的运行状态,获得了电动机、联轴器、轴承座、齿轮箱等设备的红外热图;采用Faster R-CNN算法对测量得到的旋转机械红外图像进行了学习训练和目标识别,结果表明该算法能够准确识别旋转机械部件;研究了单角度和旋转角度红外监测的识别效果,发现在相同角度下使用红外灰度图像进行训练的检测效果比使用红外伪彩色图像训练的检测效果更佳;对比了4种预训练网络对于红外目标识别的影响,采用Resnet50预训练网络的平均检测精度为0.9345,识别精度更高。

Rotating machinery is the core component of mechanical equipment and can thus cause a significant loss if it breaks down.Therefore,real-time monitoring and diagnosis of the rotating machinery is critical.Automated infrared intelligent monitoring and diagnosis is a recent development in fault diagnosis.To realize infrared intelligent monitoring and diagnosis,it is necessary to accurately identify rotating machinery components.In this study,an infrared thermal camera was used to monitor the running state of the rotating machinery and infrared images of the motor,coupling,bearing seat,gearbox,and other equipment.The Faster R-CNN algorithm was used to train the rotating-machinery infrared images and to identify the targets.The results showed that the algorithm can accurately identify rotating machinery components.The recognition effect of single-angle and rotating-angle infrared monitoring was studied.It was found that the detection effect of infrared gray images fortraining at the same angle is better than that of infrared pseudo-color images.The influence of four types of pre-training networks on infrared target recognition was compared.The average detection accuracy of the resnet50 pre-training network was 0.9345,and the recognition accuracy was higher.