Automated Defect Recognition for Additive Manufactured Parts Using Machine Perception and Visual Saliency

Area of Science:

• Materials Science and Engineering
• Non-Destructive Testing
• Computer Vision and Image Analysis

Background:

• Metal additive manufacturing (AM) processes can introduce internal defects affecting part performance.
• Nondestructive inspection (NDI) technologies are crucial for ensuring AM part quality and reliability.
• Existing methods often require prior knowledge of defect characteristics or specific part data.

Purpose of the Study:

• To develop a fully automated, 3D solution for defect recognition in metal AM parts using X-ray computed tomography (CT) scans.
• To create an algorithm that identifies internal defects without requiring a priori information about their appearance, size, or shape.
• To establish a robust method for assessing and quantifying build quality in AM components.

Main Methods:

• A machine perception framework employing multiscale, symmetric, and separable 3D convolution kernels to identify anomalous voxels (defects).
• Generation of a binary mask using statistical measures to segment material from background, accommodating arbitrary part geometries without CAD or STL files.
• Merging of adjacent anomalous voxels to form defect clusters, providing information on size, morphology, and orientation.

Main Results:

• The algorithm achieved adequate processing times, typically within minutes, using hardware acceleration (GPU support).
• Demonstrated very low false-positive rates, particularly for highly salient and larger defects.
• The developed tools can be simplified for 2D image analysis.

Conclusions:

• The proposed 3D automated defect recognition system offers a promising solution for quality control in metal AM.
• The method's reliance solely on visual cues from CT scans makes it versatile and independent of specific defect knowledge or part data.
• This technology can enhance confidence in AM build quality assessment and potentially link defect characteristics to mechanical properties like fatigue response.