Steel railway bridge fatigue damage detection using numerical models and machine learning: Mitigating influence of modeling uncertainty

Stringer-to-floor beam connections were reported as one of the most fatigue-prone details in riveted steel railway bridges. To detect stiffness degradation that results from the initiation and growth of fatigue cracks, an automated damage detection framework was proposed by the authors (Eftekhar Azam et al., 2019; Rageh et al., 2018). The proposed method relies on Proper Orthogonal Decomposition (POD) and Artificial Neural Networks (ANNs) to identify damage location and intensity under non-stationary, unknown train loads. Bridge computational models were used to simulate damage scenarios and for training the ANNs. Damage detection method efficiency and accuracy were shown to be significantly influenced by the level of modeling uncertainties (MUs). To investigate the applicability of the proposed framework to in-service bridges, a systematic analysis of the effect of MUs on the proposed POD-ANN framework was necessary. MU influence on the performance of the POD-ANN damage detection method was investigated and a new procedure for generating training data for ANNs was proposed. The procedure was based on synergizing Proper Orthogonal Modes (POMs) extracted from measured structural response and POMs calculated from the numerical model. The current study integrated numerical and field investigations. The main objective of the numerical investigation was to identify a robust damage feature independent of the level and location of assumed MUs. Results showed that Damage Location (DL) and Damage Intensity (DI) were detected with high accuracy for studied uncertainty cases; however, as expected, damage detection accuracy reduced as MU increased. A hybrid experimental-numerical approach was then implemented for the field investigation studies. This approach applied identified damage features from the numerical investigation to measurements from an in-service railway bridge to produce damage scenarios used to train the framework. MATLAB algorithms were developed that preprocessed field data and eliminated POM variations resulted from loading uncertainties. ANNs were trained and tested using the field strain estimated POMs from the hybrid approach and DL and DI results were obtained for the studied railway bridge under non-stationary, unknown train loads. These results show the promise of the POD-ANN method as a robust, real-time fatigue damage identification tool for steel railway bridges.