研究目的
To detect printing defects during the laser-powder bed fusion process using a combination of thermographic off-axis imaging and deep learning-based neural network architectures.
研究成果
The developed convolutional neural network can identify delamination defects and splatters with a very good accuracy of 96.80%. The model is small and light on computational costs, making it suitable for real-time operation on less powerful hardware. Future work includes evaluating the model's performance with a broader variety of materials, geometric shapes, and defect types.
研究不足
The convolutional neural network can only detect defects such as splatter and delamination. Other defect types such as cracks, pores, balling, and unfused powder have not been evaluated. Only one geometrical shape and H13 material were used for training and evaluation.
1：Experimental Design and Method Selection
The methodology involves using thermographic off-axis imaging as a data source and deep learning-based neural network architectures for defect detection. The network training employs k-fold cross validation and hold-out cross validation.
2：Sample Selection and Data Sources
The data set consisted of 4,314 RGB color images extracted from thermographic imaging during the L-PBF process of H13 steel specimens. Images were cropped to remove unnecessary information.
3：List of Experimental Equipment and Materials
SLM 280HL (SLM Solutions AG, Lübeck, Germany) for metal printing, PYROVIEW 640G/50 Hz/25° × 19°/compact + thermographic camera (DIAS Infrared GmbH, Dresden, Germany), and H13 steel material.
4：Experimental Procedures and Operational Workflow
Thermographic images were taken during the printing process, converted into AVI video files, and then into PNG images. These images were cropped and resized for processing. Data augmentation techniques were applied to the training set.
5：Data Analysis Methods
The convolutional neural network was trained using Keras with Tensorflow backend. Performance evaluation included balanced accuracy, sensitivity, precision, and Cohen’s Kappa score. Heatmaps were used to visually explain the class predicted by the network.
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
