Research On Recognition Method Of Stator And Rotor Surface Defects Based On Deep Learning

Stator and rotor are important parts of the engine phase regulator in variable valve timing system,which having great demand for automatic and intelligent surface quality detection.Stator and rotor have low strength and toughness,so the process of production and delivery is easy to cause defects such as bump,scratches,bumps and lack of material on the surface.Meanwhile,the defect size is small and difficult to detect due to its processing characteristics of powder metallurgy.Compared with the common way of manual visual inspection assisted by semi-automatic equipment,automatic inspection based on machine vision is more suitable for stator and rotor workpieces with rapid renewal,small production batch and various types,which can save labor costs and improve efficiency.However,the research of machine vision technology on the detecting and recognizing of stator and rotor surface defects is not mature enough,due to the rich surface texture and complex contour structure of stator and rotor,it is easy to excessively or missing detect on surface defects.The deep learning method has ability to detect stator and rotor surface defects under strong texture interference,but the detection of small defects such as scratches and bumps is not effective,detection and recognition methods for weak contrast and small scale defects are also the hot topic of the current research.therefore,working on recognition method of stator and rotor surface defects based on deep learning is worthwhile.In this paper,the surface detection of stator and rotor is studied.A recognition method of stator and rotor surface defects based on deep learning is proposed.Aiming at image acquisition of stator and rotor defects,enhancement on rough and uneven characteristics of small defects,and defect detection and recognition under complex texture background,the main work is as follows:(1)Overall scheme design of stator and rotor surface defects recognition system.Based on the analysis of the research objects and the technical requirements of its surface defects recognition,the selection and layout of image acquisition hardware are completed,the overall image acquisition procedures and process algorithms as well as software architecture of the defect recognition method are designed,flexible and convenient stator and rotor surface defects recognition systems are established,the coordination of the internal hardware is realized by using signal synchronization,normal illumination image and multi-angle illumination images were captured respectively,and these two types of images are parallelly processed by multithread.(2)Multi-angle illumination image acquisition system and correction method of position and attitude.Short lifetime of stator and rotor products makes it more suitable for the flexible and portable desktop detection device,but the initial detection signal of the device depends on manual control.Aiming at the problem of cumbersome and inefficient manual triggering,a dynamic image acquisition algorithm based on frame difference and feature moment was proposed.By using the gray change between continuous images and Hu moment feature of stator and rotor profile,the image acquisition and subsequent processing were automatically triggered when a stable object appeared in camera field of view.Manually placed stator and rotor have random position coordinates and rotation angles,which increases type and complexity of pattern recognition,therefore a matching and localization algorithm based on the characteristics of surface holes was proposed.The center of the workpiece is located by the feature of the central hole,then the angle is matched by the feature of process holes.The position and rotation correction of stator and rotor is accomplished by affine transformation.(3)Micro defects enhancement method based on improved photometric stereo.Small scratches and bumps are common surface defects of stator and rotor with soft material,which can not be effectively detected by two-dimensional images.Therefore,the surface height is reconstructed by photometric stereo,which can enhance the rough and uneven characteristics of defects.The classical photometric stereoscopic method requires the ideal conditions of parallel illumination with uniform luminance and lambert reflection.Aiming at the problem that the ideal conditions cannot be provided in stator and rotor scene,an improved photometric stereo method is proposed.Firstly,the model of near field illumination was established by using reflector sphere and reflector plate to correct the influence of luminance and illumination direction.Secondly,in the case of non-lambert reflection on metal surface,the diffuse reflection component in the image is extracted by guided filtering and Retinex,which improves the calculation accuracy of surface normal vector.The prior knowledge of hardware layout and illumination orientation is used to speed up the processing of photometric stereo.(4)Recognition method of stator and rotor surface defects based on improved RCF model.A recognition method of surface defects based on improved deep learning model was proposed to solve errors caused by complex processing texture,variable contour and scale change defects on stator and rotor surface.RCF model can extract target contour in complex environment and low contrast pixel by pixel,which has great potential in recognizing defects in complex texture and contour.To solve the problem that the RCF model could not handle the scale variations of stator and rotor defects well,a pyramid feature extraction structure was constructed in the backbone of the model to extract multi-scale and multi-level feature.To solve the problem that small defects are missing after pooling for many times in RCF model,precise location information from shallow stage and rich semantic features from deep stage are fused by skip connection,and the intermediate outputs are fully integrated by the attention mechanism to form a better final result.Aiming at the problem of lopsidedness distribution of samples caused by small defect area of stator and rotor,the loss function was optimized to improve the training effect of the model even using samples without defects.Based on the research of surface defects recognition method of stator and rotor,the control software of the stator and rotor surface defects recognition system was designed using integrated development environment.The software includes automatic detection unit,teaching unit,detection setting unit,data management unit and user management unit.With the data of the actual scene,overall running time of the proposed method and each process are tested.The results show that the efficiency of the proposed method meets the needs of practical application.The number and types of defects recognized on the unqualified workpieces were statistically analyzed.The results show that the overall accuracy and recall rate of the proposed method is96.87% and 92.25%.The recognition method of stator and rotor surface defects based on deep learning can effectively detect small surface defects.