Toward Reliable Corrosion Forecasts in Waterway Infrastructure Using LSTM-Based Models and Synthetic Sequences under Sparse Data Conditions

Reliable prediction of corrosion in hydraulic steel structures is crucial for maintenance planning and structural safety. However, practical inspection data are often sparse, irregular, and limited to a few time points per structure—challenging the applicability of conventional forecasting methods. This study presents a data-driven framework that enables long-term corrosion prediction under such constraints by combining domain-informed preprocessing, synthetic sequence generation, and deep learning. Starting from individual wall thickness measurements stored in the WaDiMe database, synthetic time series were constructed using interpolation and DTW Barycenter Averaging to simulate realistic corrosion progression. A CNN-BiLSTM model was trained on 14, 600 augmented sequences and evaluated using recursive forecasting strategies. The model reliably captured the general corrosion trend, including the characteristic exponential wall thickness loss, even with minimal input. Moreover, a transfer learning approach demonstrated successful model adaptation to new environments, such as canal structures, based on only a few additional time series. The framework thus offers a robust and generalizable solution for predictive maintenance in data-scarce settings.