This paper presents an enhanced neural network-bootstrap particle filtering algorithm to construct the complex relationship between the normalized strain relaxation indicators and the crack front profiles based on numerical simulation and experimental validation. The metamodel of normalized strain relaxation indicators and the crack front profile in welded plate joints under bending cyclic loadings is built based on the optimal regression neural network and finite element analysis. To overcome the uncertainties caused by the limited strain measurement, crack measurement, and different non-destructive techniques, this study further proposes new crack-related weight functions and combines a bootstrap particle filtering approach with an interpolation method to finely tune the metamodel and the crack prediction algorithm. As validated by the experimental results, the intelligent crack sizing approach demonstrates a potential solution for crack size monitoring through affordable strain gauges in the broad framework of digitally twinning the next-generation infrastructure.