A Lightweight Transformer with Corrosion Gating and Physical Embeddings for Pipeline Corrosion Growth Prediction

This is the authors' abstract. We don't add key points for this paper.

Pipeline corrosion critically threatens the safe operation of chemical industrial park pipeline networks, making accurate corrosion growth prediction essential for preventing catastrophic failures. Mechanistic models assume steady states, which conflict with in-service corrosion dynamics; data-driven approaches, however, presume complete datasets but frequently face missing parameters due to sensor failures and limited samples from costly inspections, increasing the risk of noise and overfitting. In this paper, the TinyTransCorrosion model was proposed, which is a lightweight Transformer-based corrosion growth prediction model specifically designed for small-sample scenarios. A cross-validation residual analysis is employed for data cleaning, while five physical embedding features are constructed to encode domain knowledge and compensate for missing parameters. A compact Transformer encoder containing only 6433 parameters was adopted, and a corrosion gating mechanism along with a classification (CLS) token was introduced to achieve efficient feature interaction. Evaluated on a real-world pipeline inspection dataset with 215 records, TinyTransCorrosion attains an R2 of 0.6411 and an MAE of 0.1608 mm, outperforming nine conventional baseline models, including Mean predictor, Linear Regression, Ridge, SVR-RBF, Random Forest, XGBoost, LSTM, MLP, and CNN. While these results are limited to a single-site dataset and require external validation on independent multi-source data, the proposed lightweight physics-guided architecture demonstrates promising predictive capability for small-sample pipeline corrosion assessment, with model error approaching the metrological limit imposed by field ultrasonic gauge accuracy. It provides an acceptable pathway for prioritizing inspection intervals and optimizing maintenance scheduling in resource-constrained industrial settings.