Metal Vibration Based On GoogLeNet Model Research On Diaphragm Defect Detection Method

In the increasing development of technology today,the electronics industry continues to grow,some small headphone audio equipment quickly occupy the market,into our lives.The study of metal diaphragm defect classification methods inside small headphone audio devices has also been widely focused on,and the direction of research has gradually changed from traditional manual detection methods to deep learning methods as a solution to the problems of low efficiency of manual detection methods,rough naked-eye classification,high detection cost,and high false detection rate of metal vibrating diaphragms in the production process.In this thesis,based on the existing convolutional neural network model,we use the migration learning method to study the defect classification method of metal vibrating diaphragms with small sample set.1.To address the problem that metal vibrating diaphragms are difficult to classify and recognize under the conditions of small data set and insufficient feature extraction in the process of recognizing film thesis peeling,holes,film particles and stain defects,this thesis uses the already trained Goog Le Net convolutional neural network model and migrates learning by adding a new Soft Max fully connected layer as the classification recognition layer.We use metal vibrating diaphragm image data to train the classification recognition layer with an accuracy of over 90%,and verify the feasibility and advantages of the method in metal vibrating diaphragm classification.2.In order to make the model have better generalization ability to unseen test samples in small sample sets,this thesis uses GA-SVM instead of Soft Max fully connected layer as the classification recognition layer of the whole network.The GASVM can find the optimal parameters c and g of the SVM faster during training,and then use the optimal parameters as the model parameters of the SVM to classify and detect metal vibration diaphragm defects,which improves the generalization ability of the whole network model and ensures better accuracy and model training time.3.For the actual project requirements,the detection method wants to be applied to the actual detection of the project,which needs to be systematized and engineered for implementation.Based on the above research method,the software and hardware selection of the metal vibrating diaphragm defect classification system is designed and implemented.The software interface for defect detection in this thesis is designed using python and QT designer,including the layout of display boxes,buttons,input boxes and adjustment boxes,and the detection algorithm and UI are logically and data connected,and finally the detection and result display is completed and the data is saved locally.In this thesis,we conducted experimental validation on the metal vibrating diaphragm image dataset provided by the project,and verified the feasibility of the migration learning method based on Goog Le Net model and the improvement method for metal vibrating diaphragm defect classification detection with a small sample set,and ensured a good accuracy and generalization capability,which has certain reference significance.The designed and implemented metal vibrating diaphragm defect detection system can also initially meet the needs of conventional metal diaphragm defect detection.