| Advanced composite materials have gradually become one of the four basic metal composite materials because of their excellent quality and performance.Especially in the field of China's aerospace industry.The amount of advanced composite materials directly reflects the development level of a developed country in its aviation industry application field.With the research and promotion of high-performance fiber-reinforced composite materials in various fields,high product manufacturing costs have become an increasingly prominent technical issue.Due to the many types of composite wire laying defects,high detection speed requirements,and traditional visual inspection metho ds It is difficult to accurately identify during high-speed operation,and missed and false detections are more serious.In view of the above problems,this article mainly completes the following four aspects:(1)Composite material defect detection algorithm based on co-occurrence matrix and improved Fisher discrimination In order to efficiently detect the laying defects of composite prepregs on-line,a method for classification of prepreg defects based on improved Fisher discrimination is proposed for defects such as wrinkles,scratches,and splits during the automatic laying process.This method uses image adaptive morphology processing to find high response regions of the image based on the Fourier transform and Hough transform to form a Gaussian distribution feature map of defect localization,and uses the corresponding position of the response peak as the defect localization position;and its bilinear interpolation For the classification feature map of size 32 * 32,calculate the eigenvalues such as the energy,entropy,and contrast of the multidire ctional co-occurrence matrix of the classification feature map.The redundant information is removed through the covariance matrix dimension reduction to obtain the 16-dimensional feature matrix.The linear discriminant matrix calculates the probability of the discriminant function to accumulate and determine the type of defect.Experimental results show that the recognition rate of defect types in this method reaches 88.3%.(2)Real-time detection algorithm for composite defects based on Yolov3: Due to the limitation of model parameters,the composite material defect detection algorithm based on the co-occurrence matrix and improved Fisher discrimination still has a large gap in deeper than the deep learning method.The Yolo-based composite material detection algorithm can increase the amount of parameters.At the same time,it learns better defect features through gradient descent,so as to solve the problem of poor appearance of traditional image features;and by establishing a data set,optimizing pre-trained models and loss functions,solving the imbalance of labeled data categories and regression branches The problem of unclear optimization direction.(3)Thinning pruning method of model BN layer based on Yolov3:The topic of this paper is based on engineering practice.In this link,under the premise of ensuring accuracy,a new model quantization pruning method is proposed.This method reduces the model parameters by reducing the model size by reducing the model size to 1/6.It is reduced to 1/3 of the original,and the speed of inference can reach 2.5 times of the original.After experimental verification,the test indicators of the model can be basically guaranteed,and the floating range is within 1%.In addition,the quantization-based model vertical integration and INT8 model are constructed and implemented,a nd the overall speed is increased by nearly 400%,which has very important practical significance.(4)Software development based on deep learning algorithms:On the basis of algorithm research:this article has cooperated to complete the corresponding algorithm detection software,which is explained in detail in Chapter 5,which is mainly divided into picture acquisition module,algorithm analysis module,communication module and d ata statistics module,which can complete normal Detection task. |
PDF Full Text Request