Abstract: This paper presents an early warning method based on a novel neural network to reliably identify the operational abnormality in expansion joints of long-span bridges. By integrating the Kolmogorov-Arnold network (KAN), temporal convolutional network (TCN), and channel attention mechanism (CAM), a TCN-CAM-KAN model is established for predicting the displacement in expansion joints. An early warning method for abnormal performance in expansion joints is developed based on the statistical characteristics of the prediction errors in the TCN-CAM-KAN model. Using long-term monitoring data from a long-span cable-stayed bridge, the capability of the TCN-CAM-KAN model in predicting displacement of expansion joints is assessed. The influence of training set length and input window size on the prediction accuracy is discussed and the effectiveness of the proposed early warning method is verified. The results indicate that the TCN-CAM-KAN model can accurately capture the complex nonlinear relationship between temperature and displacement of expansion joints, as well as the non-stationary properties of the time-dependent displacement. The prediction accuracy of the TCN-CAM-KAN model is significantly higher than that of conventional neural network models. The high prediction accuracy of the TCN-CAM-KAN model is maintained even when the training set length is short. As the input window size increases, the prediction accuracy increases first and then decreases, but the training time continues to increase. The proposed warning method can reliably detect abnormal displacement of the expansion joint above16 mm.