Interval prediction and anomaly detection for bridge expansion joint using an optimized ENN-aided temperature slow feature-displacement model

Bridge expansion joints accommodating temperature-induced longitudinal movements are vulnerable components whose failure risks structural safety. Existing temperature-displacement relationship (TDR) models exhibit limited detection reliability, hindered by inaccurate temperature field representation and inadequate prediction accuracy. This paper proposes a novel interval prediction and anomaly detection framework integrating temperature slow feature (TSF) and adaptive interval estimation. Firstly, slow feature analysis is employed to extract optimal TSFs from high-dimensional temperature data. Subsequently, a precise TDR model is established between TSFs and displacements using an optimized Elman neural network accounting for temperature hysteresis, followed by the introduction of adaptive bandwidth kernel density estimation for the first time to interval estimation and uncertainty quantification. Finally, a detection indicator is formulated using temperature-purified model errors, supported by a kernel density estimation-aided multilevel detection strategy to handle various degradation levels and nonnormal error distributions. Case validation shows that the proposed method achieves nearly 100 % detection rates for abnormal displacements over 5 mm, with reduced fluctuation range and reliable uncertainty quantification. For bilateral expansion joints, the average absolute errors are 1.057 and 0.852, and root mean square errors are 1.342 and 1.047, yielding over 10 % prediction accuracy improvement than traditional linear TDR models.