Research On Wear Particles Classification Of Wind Turbine Gearbox Based On Comprehensive Feature Extraction And Deep Learning

The wear particles generated during the operation of the gearbox of the wind turbine will adversely affect the operation of the equipment,and will cause irreparable damage to other components of the equipment as the oil circuit circulates.The concentration and type of wear particles can reflect the operating conditions of the equipment.Therefore,analyzing and studying the wear particles in the oil path of mechanical equipment is of great significance to the evaluation of the operating status of the equipment and the diagnosis of faults.In this paper,in view of the current inaccurate description of the characteristics of the wear particles of the wind turbine gearbox and the low degree of recognition,a parameter that describes the profile information of the wear particles is defined according to the characteristics of the research sample,and a method that can describe the wear particles more accurately is proposed.A comprehensive analysis method of features,on this basis,an wear particle recognition model based on deep learning algorithms is designed,which further improves the accuracy of wear particles recognition.The research results have theoretical significance and practical application value for the evaluation and analysis of the operation status of the wind turbine gearbox.The main research contents are introduced as follows:Firstly,the wear particles in the lubricant of the gearbox of the wind turbine are collected.Based on the method of digital image processing,the graying,smoothing,sharpening,binarization and related morphological operations of the wear particles are completed Using the Canny operator for edge detection,the contours of the wear particles were extracted,and a sample database of wear particles for the wind turbine gearbox containing 498 images of wear particles was established.Secondly,for the problem of inaccurate characterization of the wear particles,a parameter describing the profile information of the wear particles is proposed and defined as the radial deviation.According to the radial deviation,the shape abnormality and the profile deviation of the wear particles are specified.The features of perimeter,area,curvature,slenderness,area deviation and fractal dimension,etc.,proposed a comprehensive extraction method based on the shape and contour features of the wear particles.A random forest classification model was used to identify the type of wear particles in comparison with a single feature extraction method.The experimental results show that the comprehensive extraction method of wear particles characteristics proposed in this paper can effectively improve the accuracy of the identification of wear particles types.Finally,for the problem that the recognition rate using traditional classification methods is not high enough,an abrasive particle recognition model based on Residual Neural Network(ResNet)is designed.For comparison,the identification models based on BP neural network,Le Net-5 and Alex Net were established.Four models were used to identify and verify the experimental sample data.The accuracy,loss function and Convergence stability.The experimental results show that the ResNet-based abrasive particle recognition model can more accurately complete the classification and identification of wind turbine gearbox abrasive particles.