Deconvolution of ultrasonic signals using a convolutional neural network.

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Ultrasonic testing (UT) resolution is limited by flaw proximity and transducer characteristics.
• Increasing UT frequency improves axial resolution but reduces penetration depth due to attenuation.
• A key challenge in non-destructive testing (NDT) is balancing penetration depth and axial resolution.

Purpose of the Study:

• To enhance the compromise between penetration depth and axial resolution in ultrasonic testing.
• To develop a method for separating overlapping echoes in time traces for improved flaw detection.

Main Methods:

• Utilized a convolutional neural network (CNN) to analyze ultrasonic time traces.
• Trained the CNN using simulated ultrasonic data.
• Validated the framework experimentally by detecting flat-bottomed holes in an aluminum block.

Main Results:

• The CNN effectively separated overlapping echoes, enabling accurate time-of-flight and amplitude estimation.
• The developed framework demonstrated improved axial resolution and penetration depth capabilities.
• Experimental validation confirmed the detection of defects at depths as small as a quarter wavelength (λ/4).

Conclusions:

• Convolutional neural networks offer a promising approach to overcome the resolution-penetration trade-off in ultrasonic testing.
• Simulated data training is an effective strategy for developing robust NDT AI models.
• The proposed method advances the capabilities of ultrasonic non-destructive testing for material characterization.