Accurate structural behavior interpretation via finite element models is often disrupted by uncertainties, while data-driven approaches can struggle with long datasets, complex fluctuations, and the omission of essential spatio-temporal features. Additionally, these methods are limited by their reliance on interpolative predictions. This paper introduces a novel, model-free approach that integrates Moving Principal Component Analysis (MPCA), bidirectional gated recurrent units (biGRU), and attention mechanisms (AM) within an encoder–decoder (ED) architecture. MPCA reduces dimensional complexity, extracts spatial features, and consolidates them into new time-series data for subsequent analysis. The biGRU module captures past and future dependencies, while AM emphasizes most relevant information. Validated on a full-scale pedestrian bridge dataset, the presented MPCA–biGRU–AM model converges 19% faster than MPCA–GRU and reduces anomaly detection lag by 46–78%. Although its per-step processing time (8 ms) slightly exceeds that of MPCA–GRU (3 ms), the model demonstrates greater robustness across diverse damage scenarios. These results highlight its potential for real-time structural health monitoring by effectively capturing spatio-temporal patterns with computational efficiency.

Self-sensing concrete (SSC) can support structural health monitoring by turning the bulk electrical response of cementitious material into a sensing channel. This paper evaluates simple, training-free resistivity-derived features under monotonic destructive tests (MDT) and ramp--hold--ramp monitoring (NMT-RHR). Graphite--milled carbon fibre concrete cubes were tested while resistance, strain, and force were recorded. The electrical response was processed into fractional resistivity change ($\kappa$), its time-rate ($\beta_t$), and supporting strain-normalised variants. Low-load segments were treated as conservative internal references for feature comparison rather than as verified damage-free states. Across the retained MDT specimens, $\kappa$ provided the clearest sustained deviation from the initial reference segment, while $\beta_t$ highlighted abrupt late-stage transitions. Across NMT hold segments, $\kappa$ showed the strongest sustained response as load severity increased, whereas $\beta_t$ behaved as a more transient companion indicator. The strain-normalised variants were less stable and are best treated as secondary comparison features. These results indicate that sustained and rate-sensitive electrical features provide complementary information: $\kappa$ is more suitable for tracking persistent offsets, while $\beta_t$ is more suitable for identifying short transitions. Overall, the study provides a comparative basis for selecting simple electrical features under controlled compression loading, with the reported detections interpreted as reference-relative changes in the electrical response.