xLSTM-m for Multivariate Structural Response Classification in Bridge Monitoring via Masked Mean Pooling

Multivariate structural-response classification supports bridge monitoring by integrating synchronized measurements from multiple sensors and assigning an observation window to a predefined structural or loading condition. The main technical challenge is to compress temporally and spatially distributed sensor information into a fixed-length representation without losing transient or localized condition-sensitive features. However, previous studies have mainly focused on backbone architecture design, whereas the influence of sequence-level aggregation has rarely been isolated under a controlled backbone and training configuration. In this study, xLSTM-m is proposed as a modified variant of the extended Long Short-Term Memory (xLSTM) network for multivariate bridge-response classification. The conventional last-hidden-state readout is replaced with masked mean pooling over all valid hidden states, while the hybrid sLSTM/mLSTM backbone and the remaining training configuration are kept unchanged. This controlled design isolates the sequence-level aggregation strategy as the only architectural variable. A comparative evaluation is conducted using eight neural models. These models include xLSTM-m, the baseline xLSTM, and six Transformer-based architectures. Three bridge-monitoring datasets are considered: an FBG strain-response dataset and a PCB Piezotronics acceleration-response dataset acquired from a laboratory-scale cable-stayed bridge model, together with the field-scale Z24-9Setup acceleration benchmark. A strict five-fold cross-validation protocol is adopted. The proposed xLSTM-m ranks first on all three datasets. It achieves a mean accuracy of 95.29%, a mean F1-score of 95.57%, and the lowest inter-fold standard deviation of 0.84. Relative to the strongest Transformer baseline for each dataset, the corresponding accuracy gains are 6.30 percentage points for FBG, 3.66 percentage points for PCB, and 31.76 percentage points for Z24-9Setup. These results indicate that, for the evaluated bridge strain and acceleration datasets, sequence-level aggregation substantially affects both classification accuracy and inter-fold stability. Further validation using quasi-static responses, environmental variables, other structural systems, and unseen operating conditions is required before broader applicability across the SHM domain can be established.