Pipeline monitoring data recovery using novel deep learning models: an engineering case study

Pipeline monitoring frequently encounters missing data, leading to incomplete evaluation and hindering a comprehensive assessment of the pipeline's structural health. To address this issue, this study proposes a novel PDO-BiGRU-GAN model for missing data recovery. The model integrates three components: the prairie dog optimization algorithm (PDO) for hyperparameter tuning, the bidirectional gated recurrent unit (BiGRU) for effective temporal feature extraction, and the generative adversarial network (GAN) for data generation and completion. A comprehensive monitoring database was established using field data from an open-source pipeline project. The contributions of individual modules to the overall performance were evaluated via hyperparameter sensitivity analysis and ablation studies. The impact of missing data ratio and the number of missing sensors on the model's recovery performance was analyzed. In addition, the proposed model was compared with eight existing mainstream deep learning models. The results show that each component of the PDO-BiGRU-GAN significantly enhances overall performance. The model achieves strong recovery accuracy across various missing data scenarios, with the R² consistently exceeding 0.93. Moreover, the model performs optimally when the missing data ratio is below 20/24. Compared to other models, PDO-BiGRU-GAN achieves the highest R² and the lowest error metrics (MSE, RMSE, MAPE, MAE). In terms of computational efficiency, the model requires slightly more processing time than simpler models but is faster than more complex models. Overall, the proposed model provides a robust and scalable solution for pipeline monitoring data recovery, advancing intelligent pipeline health assessment and supporting the development of infrastructure safety management and smart monitoring technologies.