Autoencoder-based detection of near-surface defects in ultrasonic testing.

Area of Science:

• Non-destructive testing
• Materials science
• Signal processing

Background:

• Defect detection in pulse-echo ultrasonic testing (UT) is difficult in the 'dead zone' due to transducer ringing.
• Conventional time-gate methods fail to detect defects within this dead zone.

Purpose of the Study:

• To propose an autoencoder-based end-to-end UT method for detecting defects in the transducer dead zone.
• To enhance autoencoder performance with limited data using a novel two-step training procedure.

Main Methods:

• An autoencoder is designed to predict normal ultrasonic signal behavior, including disturbances.
• A two-step training involves initial training on normal signals and re-training on autoencoder-identified pseudo-normal samples.
• The method was demonstrated using B-scan inspections of aluminum blocks with near-surface defects.

Main Results:

• The autoencoder successfully identifies subtle deviations caused by defects.
• The two-step training procedure creates an adaptive model for processing new ultrasonic signals.
• The proposed method significantly outperforms conventional gate-based approaches in detecting near-surface defects.

Conclusions:

• The autoencoder-based UT method effectively detects defects in the dead zone.
• The adaptive autoencoder model improves defect detection accuracy and localization.
• This approach offers a superior alternative to traditional methods for challenging UT defect identification.