| Fiber optic gyroscope is an inertial sensor with low cost,small size and simple structure.Fiber optic gyroscopes are widely used in inertial navigation devices,and fiber optic rings are its core components.In the actual winding process of the fiber optic ring,problems such as "stacking(overlap)","collapse",and "gap" are likely to occur.The fiber itself may also have cracks,breakage,crushing,etc.,and the defective fiber ring will affect the measurement accuracy of the fiber optic gyroscope.In the current fiber-optic ring winding process,the method of fiber defect detection and fiber winding defect monitoring is mainly based on manual monitoring.With the maturity of technologies such as image processing,machine learning,and deep learning,the use of machine vision to identify and classify fiber defects is the development trend of defect detection technology.In this paper,fiber defect detection and fiber winding defect monitoring technology are taken as research objects.It is difficult to detect the arrangement defects and fiber defects occurring in the fiber winding system for manual detection.Combined with the development status and shortcomings of the current winding defect monitoring technology,machine vision technology,image processing,machine learning and other methods are used to study the fiber winding defect monitoring and fiber defect detection.The following is the specific research content:Firstly,the convolutional neural network is applied to the fiber defect detection technology,and a SVM image classification method based on convolution feature is proposed.The method uses the ResNet50 model to extract the 2048-dimensional convolution feature of the fiber,and uses the PCA method to reduce the dimension according to the feature retention ratio.The SVM parameters are optimized by PSO and the images are classified using SVM.This method uses convolutional neural networks to extract features,avoiding the excessive dependence of traditional feature extraction methods on prior knowledge and experience.The SVM is more suitable for small sample classification comparing with the multi-layer perceptron of neural network for image classification,which avoids the over-fitting problem caused by insufficient sample in the convolutional neural network.Simulation experiments verify the superiority of this method in identifying minor defects.Secondly,for the fiber defect detection technology,it is proposed to improve the LBP feature extraction method by using Gabor filtering and Sobel filtering,and improve the LBP feature to extract the micro defect features of the fiber.A method of combining the HOG feature of the fiber with the improved LBP feature into a combined feature for SVM classification is proposed.The combined feature includes the gradient feature extracted by HOG and the texture feature extracted by LBP,and further exploits the defect characteristics of the fiber.Simulation experiments verify that the use of joint features has higher accuracy,and improved LBP features can better identify minor defects.Finally,for the fiber winding defect monitoring technology,an image classification method is proposed to replace the "overlap" and "too large gap" defects in the traditional measurement identification winding.At the same time,the sub-pixel measurement is used to assist the image classification to determine the "over-spacing" and "uneven pitch" states that occur during the winding.Compared with the traditional single measurement method,there is no need to change the determination condition of the arrangement state according to the fiber ring parameters.This method has low requirements on the accuracy of the measuring equipment and the preprocessing requirements of the image,which improves the measurement accuracy and reduces the measurement parameters.Simulation experiments verify the feasibility of this method. |
