Abstract Diagnosis of underground drainage pipeline siltation is an important means of maintaining and operating buried pipeline geographical infrastructure and preventing and reducing disasters. Deep learning prediction algorithms are widely adopted in intelligent prediction and diagnosis examples related to drainage pipeline operation and maintenance. However, existing prediction neural network algorithms have problems such as being unsuitable for multivariate data sequence analysis, extracting single features, and output results that are prone to partial distortion. This study proposes an intelligent diagnosis model based on the fusion of an improved Kolmogorov-Arnold network (IKAN) module and a Transformer network to address the problem of insufficient extraction of nonlinear features in the diagnosis of siltation in urban drainage pipelines. The model first adopts a IKAN layer constructed using an adaptive grid division and spline interpolation mechanism to convert the nonlinear features of the flow rate and flow velocity data at the inlet and outlet of the pipeline, thereby overcoming the limitations of traditional fixed activation functions that rely too heavily on linear weight matrices and enhancing the model’s ability to express complex data patterns. Subsequently, one-dimensional convolution and fully connected layers are adopted to extract local features, and the Tokenization module is adopted to map the data to a high-dimensional token space. Finally, the multi-layer, multi-head self-attention Transformer network is input to achieve deep modeling of spatiotemporal dependencies, and dual-task learning is adopted to perform regression and classification simultaneously. Compared with other prediction models, the model proposed in this paper is significantly superior to other algorithms in terms of the diagnostic accuracy of the thickness/length of siltation in the pipeline. The experimental results show that the model obtains the best performance when the time step is set to 15 s. The intelligent diagnostic model proposed in this paper has strong robustness and prediction accuracy, achieving 0.282, 0.248 and 0.185 in the evaluation metrics RMSE, MSE and MAPE, respectively, when the proportion of noisy sample data is less than 50%.