Research On Stator Coil Defect Detection Algorithm Based On Deep Learning

The stator is an important part of basic electric equipment such as generators and motors.The quality of the coil has a direct impact on the performance of the corresponding equipment.and the quality of the stator is closely related to the state of the machinery and equipment on the production line.The status of the production line machinery and equipment can be reflected by detecting the coil defect,and the winding equipment fault diagnosis can be performed when the machine is backtracking,thereby correcting the winding equipment and reducing the rate of defective products.This thesis applies machine vision technology to coil defect detection system.We proposes a deep learning-based coil defect detection algorithm,which overcomes the shortcomings of manual detection and traditional machine learning feature engineering,and can implement learning activly and extracts defect features automatically.Aiming at the problem of insufficient sample data,a model-based transfer learning method was proposed to overcome the difficulty of model training with small samples.Firstly,according to the research object of this thesis,combined with the actual environment,an overall detection scheme was designed.In order to highlight the defect feature information,the coil area needs to be divided into blocks becaue that the type of coil defects is relatively small overall At first,a coil region is initially extracted,and a coil region extraction algorithm based on the RGB model is proposed based on the color characteristics of the coil.At this time,the extracted coil area has noise problems such as holes,and image preprocessing algorithms such as image binarization,morphological operation are used to eliminate noise to obtain a more complete coil area.Next,a coil area block algorithm is proposed.The matrix area is used to divide the coil area into rectangular blocks of corresponding size for storage.We classify and label the saved rectangular blocks to make a data set required for network model training.Secondly,the hierarchical structure of the convolutional neural network is studied in depth.In order to make the model more stable during training,the optimization method in the training process is studied.In this thesis,the VGG16 model is selected as the classifier and the training required parameters.The original data set has fewer images,and it is easy to cause overfitting when it is directly trained in the network model.The image enhancement method is used to expand the data set and then train the model.After training,the accuracy rate and loss function curve are obtained.After a certain number of iterations,the accuracy rate of the model is above 95%.The single coil instance is detected by the model,which proves that the model detection performance meets the needs and can provide a reference for equipment correction.Finally,the training method in the case of small samples is studied.According to the implementation of transfer learning,the model-based transfer learning method is selected to fine-tune the VGG16 pre-training model in the source domain to the target domain,that is,to freeze the "bottleneck layer" of the model,and retrain the Softmax classifier layer with coil samples.After training,the accuracy and loss function curves of the model are obtained.The accuracy of the front,side and slope are above 95%.The experiment shows that transfer learning can achieve good training results in the case of small samples.According to the experiment in this thesis,when the number of samples collected is sufficient,the model obtained by using the non-transfer learning method is better than the use of transfer learning to train a part of the hierarchy;when the number of samples collected is small,you can consider using the transfer learning method to train the network model.