Time-series anomaly detection and interval-based risk identification for station roof slab displacement based on multi-model consistency fusion

To address the problems of strong noise interference, lack of anomaly labels, unstable recognition by a single model, and limited engineering applicability of point-level anomaly prompts in construction-stage roof slab displacement monitoring data from shallow-buried, large-span integrated transportation hub stations, this study proposes a time-series anomaly detection and interval-based risk identification method for roof slab displacement based on multi-model consistency fusion. Taking the newly built station project of the Huangmugang Comprehensive Transportation Hub in Shenzhen as the engineering background, the roof slab displacement time-series data obtained from machine-vision monitoring are first cleaned, time-aligned, and normalized. Then, autoregressive integrated moving average (ARIMA), Isolation Forest (IF), local outlier factor (LOF), and long short-term memory autoencoder (LSTM-AE) are jointly employed to extract anomaly scores from the perspectives of prediction residuals, isolation degree, local density, and reconstruction error. Subsequently, through quantile normalization, a multi-model consistency index, and fused anomaly probability, the outputs of different models are mapped onto a unified scale, thereby reducing the influence of single-model bias and random noise on anomaly identification. Finally, by incorporating constraints on minimum duration and maximum merging gap, discrete anomalous points are aggregated into anomalous intervals with start time, end time, duration, and intensity attributes, followed by risk classification and spatial correlation analysis. Independent validation on the Huangmugang project and on multi-point monitoring data of Shanghai Metro Line 11 shows that different single models have complementary recognition capabilities for anomaly patterns such as short-term abrupt changes, local outliers, and persistent pattern changes. The multi-model fusion results can stably identify anomalous intervals of roof slab displacement under noisy conditions and reflect the staged characteristics and spatial correlations of structural responses induced by construction disturbances. The proposed method transforms high-frequency roof slab displacement monitoring data from point-level anomaly prompts into interval-based risk events, providing technical support for the in-depth utilization of structural deformation monitoring data and refined risk identification during the construction of complex integrated transportation hubs.