Towards using convolutional neural network to locate, identify and size defects in phased array ultrasonic testing.

Area of Science:

• Non-destructive testing
• Machine learning applications
• Ultrasonic imaging

Background:

• Phased Array Ultrasonic Testing (PAUT) is crucial for flaw detection but often slow for high-speed manufacturing.
• Current PAUT data acquisition requires multiple emissions, limiting inspection speed.
• Faster R-CNN offers potential for accelerated image recognition in NDT.

Purpose of the Study:

• To develop and evaluate a Faster R-CNN model for rapid identification, localization, and sizing of defects in PAUT.
• To enable high-speed inspection by using a single plane wave insonification.
• To compare the performance of the model against traditional methods.

Main Methods:

• Utilized Faster R-CNN for defect detection in an immersed test specimen.
• Trained the model on simulated data (finite element simulations) and a small experimental dataset.
• Employed single plane wave insonification for fast data acquisition.
• Assessed results using Intersection over Union (IoU) metrics.

Main Results:

• The model achieved high detection rates for flat bottom holes (FBH) (87% in simulation, 70% in experiments) at a 40% IoU threshold.
• Simulations showed accurate detection of specimen thickness and FBH down to 0.56 wavelength depth.
• Experimental results confirmed reliable specimen identification and FBH detection.

Conclusions:

• Faster R-CNN with single plane wave insonification enables significantly faster PAUT data acquisition.
• The proposed method shows promise for real-time, high-speed industrial inspection applications.
• Further development may improve detection rates for side-drilled holes (SDH).