研究目的
To minimize the probability of human error in shearography inspection by developing a convolutional neural network for binary classification of defects in shearography images.
研究成果
Convolutional neural networks can learn from shearography fringe patterns, with architecture B achieving 79% accuracy in binary defect classification. The combination of double filter layers and dropout layers is effective. Future work should involve multi-classification and parameter engineering for better defect characterization.
研究不足
The dataset is small (294 samples), which may limit model performance. No grid search was performed to explore all possible hyperparameter combinations, and the study focused only on binary classification, not multi-classification. Accuracy achieved (79%) is considered low, potentially due to the small dataset.
1：Experimental Design and Method Selection:
The study involved training two convolutional neural network architectures (A and B) based on VGG characteristics for binary classification of shearography images into major and minor defects. The methodology included data augmentation, dropout layers, and hyperparameter tuning to avoid overfitting.
2：Sample Selection and Data Sources:
The dataset consisted of 294 shearography test samples from pipes repaired with glass fiber patches, with 168 major defects and 126 minor or no defects. Images were obtained from shearography tests on pipeline samples with artificial defects.
3：List of Experimental Equipment and Materials:
A shearography setup with an expanded laser source, camera, and shearing device; pipeline samples with composite repairs (NRI Syntho-Glass, WTR Technowrap 2K, Furmanite FurmaCarbon, Clock Spring); artificial defects (machined slots, blind holes).
4：Experimental Procedures and Operational Workflow:
Shearography phase difference maps were captured before and after load application. Images were preprocessed, augmented (scaling, shear, rotation), and used to train CNNs in Python with TensorFlow and Keras on a GPU. Training involved batches of images, early stopping, and validation.
5：Data Analysis Methods:
Accuracy and loss metrics were computed for training and validation sets. Hyperparameters (number of detectors, stride, dropout percentage, neurons in dense layers) were varied and evaluated.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
