Dynamic deflection monitoring with distributed photoelectric deflectometers (PDs) is pivotal to lightweight bridge structural health monitoring systems (SHMS). With few sensors yet theoretically unlimited monitoring markers, such systems maximize coverage under constrained investment; however, minimal redundancy makes them acutely vulnerable to sensor-level anomalies This vulnerability is amplified in small- and medium-span bridges where high-frequency, multi-channel streams must be fused in real time, demanding anomaly detection (AD) that is both robust to environmental variation and computationally efficient with minimal preprocessing. To address this, we propose SGFormer, a transformer-based AD framework featuring similarity-grouped fusion and dynamic prototype learning for multi-sensor deflection signals. A group-level feedforward module organizes channels with aligned temporal embeddings to exploit strong intra-span correlations while preserving cross-span differences, reducing redundancy and stabilizing fusion. Self-updating prototypes, learned directly from lightly preprocessed raw signals, encode operational and environmental effects, thereby avoiding heavy handcrafted calibration. We evaluate SGFormer using representative PD anomalies that reflect the essential characteristics of in-service bridge monitoring systems in lightweight SHMS. Experiments show that SGFormer consistently outperforms seven state-of-the-art baselines in unsupervised and zero-shot settings, achieving average F1-scores of 91.49% and 89.05%, with robustness across temporal granularities.
Zhu et al. (Sat,) studied this question.