Reliable real-time structural health monitoring (SHM) for high-value composite assets requires not only accurate full-field state reconstruction from sparse sensors, but also spatially resolved uncertainty to support risk-aware decisions in digital-twin pipelines. [1] This work presents an integrated framework that reconstructs full-field strain distributions from limited strain gauge measurements while explicitly separating aleatoric (data-inherent) and epistemic (model) uncertainties in full-field form. [2] High-dimensional DIC strain fields are first compressed using principal component analysis (PCA), retaining eight dominant modes that preserve 95% of variance. A Bayesian neural network (BNN) then maps 12 strain-gauge inputs to the reduced PCA coefficients, enabling rapid inference and uncertainty-aware prediction under noisy experimental conditions and crack-tip strain singularities. [3] A key novelty is a mode-aware probabilistic formulation: during pre-training, the BNN learns mode-wise predictive variances via a heteroscedastic Gaussian likelihood, [4] yielding calibrated aleatoric uncertainty per PCA mode. These learned, data-driven variances are subsequently used as likelihood scales in Hamiltonian Monte Carlo (HMC) posterior sampling, replacing common single-variance heuristics and improving posterior consistency across modes with unequal importance. The resulting posterior predictive distribution provides (i) full-field mean strain reconstruction and (ii) complementary full-field aleatoric and epistemic uncertainty maps for local diagnosis. The framework is validated on cyclic four-point bending tests of CFRP specimens with varying crack lengths (1–15 mm), achieving accurate field reconstruction (R² > 0.9) and practical real-time performance. Joint interpretation of the two uncertainty fields enables localization and attribution of low-confidence regions to either measurement/data issues or model insufficiency, advancing trustworthy SHM toward deployable, uncertainty-aware digital twins.