Weld Width Measurement and Defect Detection Method for Desiccant Cartridges Based on Machine Vision.

In industrial automated production, reliable inspection of welded desiccant cartridges is challenging because polypropylene welds exhibit irregular boundaries, low-contrast defects, and non-uniform surface reflections. This study proposes a machine vision-based method for integrated weld width measurement and multi-type weld defect detection. A cubic B-spline fitting strategy is first employed to extract the weld boundaries and suppress interference from non-weld regions. The weld width is then measured along the local normal direction using an orthogonal projection method, with the minimum width used to identify excessively narrow welds and weld breakage. For pore and burn-through detection, distance transform-based watershed segmentation is combined with grayscale polarity analysis to separate true defects from pseudo-defects. The polarity thresholds are determined using an independent calibration batch and further verified on a separate production batch before evaluation on an independent 500-sample test set. Experimental results show that the proposed method achieves an average IoU of 96.7% for weld region extraction. The average and minimum width absolute errors are 0.0509 mm and 0.0438 mm, respectively, with relative errors of 1.69% and 1.72%. Compared with four representative geometric width measurement methods, the proposed method achieves the lowest measurement errors while requiring only 1.84 ms for width measurement. On the independent 500-sample test set, the proposed system achieves an overall defect detection accuracy of 97.20% and a Macro-F1 score of 97.82%. The average end-to-end processing time is 55.48 ms per sample, demonstrating the potential of the proposed method for real-time weld quality inspection of desiccant cartridges under the evaluated industrial imaging conditions.