Abstract
This paper introduces a novel differential TMR-ACFM probe integrated with deep learning for crack detection and evaluation. The differential design effectively mitigates the lift-off effect and external noise, thereby enhancing detection performance without increasing costs. A miniature TMR was designed and fabricated for the probe. Two TMR units were integrated in an area of 175 × 200 microns, and two dies formed the differential structure of the Wheatstone bridge. Experimental results indicate that, in comparison to conventional probes, the quality factor of the differential probe is improved by more than an order of magnitude, and the signal-to-noise ratio is enhanced by over 3 dB. Additionally, a CNN + CBAM network is developed and trained on experimental data to achieve high-precision evaluation of crack dimensions. For cracks measuring 10-30 mm in length, 2-6 mm in depth, and 0.25-1.25 mm in width, the relative errors in the predicted dimensions are 0.201%, 0.709%, and 7.224%, respectively. These results underscore the significant potential of the proposed approach for quantitative crack detection.