A transfer learning-based approach to fatigue life prediction of corroded bimetallic steel bars using small samples

•Transfer learning-based framework was proposed for fatigue life prediction of corroded bimetallic steel bars.•Transfer models showed high prediction accuracy compared with baseline models.•Convolutional and backpropagation neural networks were compared.•Influence of critical parameters on fatigue life were investigated. The low-cycle fatigue life of reinforced steel bars is usually determined by time-consuming tests. To investigate the fatigue behaviour of corroded steel bars, corrosion tests need to be conducted before low-cycle fatigue tests, which further increases time and costs. Although typical machine learning can provide an efficient solution to fatigue life prediction, the inherent need for sufficient training data remains unaddressed. By leveraging the existing test data of other metallic bars, a cost-saving transfer learning-based framework that was able to predict the fatigue life of corroded steel bars made of new types of metallic material using small samples was proposed in this study. A source dataset of eleven types of metallic bars with 450 samples and a target dataset of bimetallic steel bars (BSBs) with 54 samples were collected. Convolutional neural networks (CNN) and backpropagation (BP) neural networks were employed and compared for fatigue life prediction. Two case studies were conducted to validate the proposed framework. Compared with baseline models directly built using the target dataset without transferring the deep features learned by source models, transfer models performed better in predicting the fatigue life of BSBs with much higher accuracy. To be specific, the value of coefficient of determination (R2) for fatigue life prediction of the baseline model was negative, indicating that the target dataset was insufficient to train an isolated model with good accuracy. In contrast, by transferring deep features extracted from the source dataset and stored in source models, an R2 value of 0.883 on the testing set of the target dataset was achieved by transfer models. Additionally, the constructed CNN models offered overall better fatigue life predictions for both the source and target datasets compared with the constructed two-layered BP models. Precisely, for the CNN models, the values of R2 on the target testing sets in Cases I and II were 0.883 and 0.866, respectively, both of which were higher than the corresponding values of 0.812 and 0.798 achieved by the BP models. Furthermore, based on the constructed models, the effect