Magnetic Tile Defect Detection Based On Convolution Neural Network

Since the 21st century,with the progress of science and technology and the slogan "Science and technology is the first productive force",the existence of motor has been known by more and more people,and has been widely used in all aspects of production and life.Permanent magnet motor has become the "backbone" of the motor because of its simple structure,small size but high efficiency and other advantages.The quality of magnetic tile,the main working component,will directly affect the working performance and service life of permanent magnet motor,and then affect the quality of production and life,so the fault defect detection of magnetic tile is particularly important.The common detection method is still based on visual inspection of the detection personnel.Due to human visual fatigue and other factors,it is easy to missed detection or even error detection.In this paper,a magnetic tile defect detection method based on convolutional neural network is proposed,which can meet the requirements of large number and heavy task magnetic tile defect detection and classification.In this paper,five defects of magnetic tile produced when leaving the factory are tested,and the completed work includes the following points:(1)Generation and enhancement of data sets.To solve the problem of uniform distribution in the original magnetic tile data set,an improved DCGAN generative adversarial network is designed in this paper.The residual structure is used to replace the original convolution layer,so as to avoid gradient disappearance or explosion and network degradation.The image quality experiment verifies that the image generated by the improved network can meet the requirements of magnetic tile defect detection.The data set required for subsequent experiments was expanded and constructed by image size transformation.(2)Selection and improvement of main feature network.In many convolutional neural networks,the above data sets are used for training and classification,so as to find the most suitable network for magnetic tile defect detection is ResNet-18 network,but the accuracy is not satisfactory.In this paper,improvements are made to the ResNet-18 network,including the use of a multi-channel in-branch structure and the use of ECA modules for channel feature extraction,and the use of Dropout to randomly deactivate neurons at the full-connection layer.(3)Experimental verification analysis.The image quality evaluation index is used to evaluate the generated image.Seek the most suitable network model for magnetic tile defects;Verify the feasibility of each improved network part by experiment;Verify the performance of the improved overall network;Experiment on anti-noise performance of network model.Through experimental verification,comparison experiment of data sets,comparison experiment of network model,and comparison with different methods used by other scholars,it can be seen that the improved ResNet-18 network adopted in this paper meets the requirements of magnetic tile defect detection task and provides a new idea and method for magnetic tile defect detection.Figure 42 Table 12 Reference 65...