Machine learning-based bridge cable damage detection under stochastic effects of corrosion and fire

This paper proposes a novel machine learning-based cable damage detection model to investigate the upper and lower bounds of bridges' cable damage degrees under the effects of corrosion and fire. In the proposed approach, the surrogate model for bridge cable damage detection under stochastic effects of corrosion and fire was established by combining machine learning and finite-element analysis to estimate the remaining life of cables. Then the accuracy and generalization performance of three typical machine learning methods for cable damage prediction are compared, such as Back Propagation neural network(BPNN), Radial Basis Function neural network (RBFNN) and Least Square-Support Vector Machine (LS-SVM). It is conducted that LS-SVM owns better prediction accuracy for cable damage under the coupling effects of corrosion and fire than the others. Additionally, the LS-SVM surrogate model combined with stochastic analysis and time-dependent deterioration model of steel wires under corrosion and fire is used to obtain the upper and lower bounds of cable damage under coupling effect of corrosion and fire. The proposed surrogate model can assist management in diagnosing and evaluating cable damage more quickly, efficiently, and flexibly once the real-time monitoring data is obtained. In addition, the surrogate model can guide bridge maintenance in advance.

Automated summary

This paper proposes a novel machine learning-based cable damage detection model to investigate the upper and lower bounds of bridges' cable damage degrees under the effects of corrosion and fire. The proposed surrogate model combines machine learning and finite-element analysis to estimate the remaining life of cables. LS-SVM shows higher prediction accuracy for cable damage under the coupling effects of corrosion and fire. The proposed surrogate model can assist management in diagnosing and evaluating cable damage more quickly, efficiently, and flexibly once real-time monitoring data is obtained.