Research On Visual Measurement Of Gear Pitting Based On Tree Cycle Generator And Attention Deeplabv3

With the progress of science and technology and the increase of social demand for production,the mechanical equipment continues to develop towards high efficiency,high speed and high precision.Gearbox is a key component of transmission torque in the transmission system of mechanical equipment.Once a fault occurs,it will be related to the normal operation of the whole mechanical equipment.Gear pitting is one of the most common failure forms of gear,pitting area ratio can be used to judge the state of pitting fault.Therefore,the quantitative detection of pitting area ratio is of great significance to ensure reliable operation of machinery and reduce maintenance cost.The traditional gear pitting detection technology adopts manual observation method,but this method has some problems such as low efficiency and accuracy,forced shutdown and high expert cost.This paper explores a visual detection method based on tree cycle generator and Attention Deeplabv3 based on the above problems,which is used to realize non-contact quantitative detection of gear pitting area ratio.The main research contents of this paper are as follows:1.Design of image acquisition device for gear pitting: The demand and existing problems under actual working conditions are analyzed.and a pitting image acquisition device based on gear contact fatigue strength test machine is designed.The device is integrated in the FZG standard test bench and the imaging system using high-speed CCD By moving the platform,the position and angle of the lens are adjusted to the optimal position,that is,the camera is perpendicular to the effective tooth surface of the gear to achieve the purpose of tunable acquisition of the complete tooth surface image by the mobile platform in the actual operation process.2.Research on sample augmentation technology of gear pitting: In order to solve the problems of high cost of pitting collection and small sample,a tree cycle generative adversarial network(TCGAN)was proposed.The generative network uses tree generator to generate high-quality target samples of various styles from the source domain samples,and amplifies the differences between any two branches through maximum diversity loss(Md Loss).The quality and diversity of acquired images obtained by TCGAN were evaluated by objective indexes,and the experimental results show that TCGAN has better generation performance and can effectively augment the gear pitting data set.3.Research on gear pitting recognition and segmentation: In this paper,a segmentation network named Attention Deeplabv3 with attention mechanism is proposed.By embedding channel attention module and spatial attention module in Deeplabv3+,the feature segmentation ability of the model for small and irregular objects is enhanced,and the convergence speed of the model is accelerated.The model is applied to the segmentation of gear pitting and tooth surface and the calculation of gear pitting area ratio.Experimental results show that Attention Deeplabv3 has better segmentation performance and higher measurement accuracy.