Research On Complex Metallographic Structure Recognition And Grading Methodbased On Deep Learning

In the field of materials science and engineering,with the development of data-driven materials informatics and computational materials science,intelligent and high-throughput image recognition and characterization methods represented by deep learning have become hot topics in domestic and foreign research.In response to the problem that traditional image processing techniques are difficult to efficiently and automatically identify and rate complex metallographic structures,a sample dataset for complex metallographic structure identification and rating is established based on the heat-resistant steel metallographic inspection images used in thermal power units.Based on deep learning,a convolutional neural network model is constructed to study the intelligent identification and rating method for complex metallographic structures.The main research content and conclusions are as follows:(1)A large number of metallographic inspection images of heat-resistant steel used in thermal power units were collected and manually identified,classified,and rated according to relevant standards.A dataset for identifying complex metallographic structures,a dataset for rating ferrite + pearlite structure spheroidization and bainite structure grain size,were established.The dataset was randomly split into a training set and a validation set in a ratio of approximately 9:1 to provide training samples and performance evaluation baseline data for the subsequent construction of deep learning models for complex metallographic structure identification and rating.(2)Based on the Conv Ne Xt-T convolutional neural network model,a fully connected layer was added and the model was improved using the Dropout method to alleviate overfitting and improve generalization.A recognition model and two rating models were built for complex metallographic structures.Data preprocessing,data augmentation,label smoothing,and transfer learning methods were used to improve the training effectiveness and performance of the models.The performance of the models was evaluated comprehensively using different indicators such as accuracy,precision,recall,specificity,and F1-score based on the confusion matrix.(3)The mean values of the five evaluation indicators for the complex metallographic structure recognition model are 98.1% for accuracy,95.1% for precision,94.9% for recall,98.9% for specificity,and 95.0% for F1-score.The mean values of the five evaluation indicators for the ferrite + pearlite structure rating model are 98.7% for accuracy,97.3% for precision,97.2% for recall,99.1% for specificity,and 97.2% for F1-score.The mean values of the five evaluation indicators for the bainite structure rating model are all 96.6%.The recognition model and rating models were combined into a joint model that first performs intelligent identification of complex metallographic structures and then performs intelligent rating of ferrite + pearlite and bainite structures based on the identification results.The average accuracy for rating ferrite + pearlite structures is 97.5% and for bainite structures is96.0%.The processing speed of the joint model for images was approximately 87 images per second,as verified by experiments.The research results show that the constructed joint model can accurately identify and rate complex metallographic structures,providing a new method for efficient and intelligent identification and rating of complex metallographic structures.