Machine Learning technique for the diagnosis of environmental degradation in a steel structure

In recent years, monitoring techniques for knowledge of structural safety have become increasingly important. Machine Learning techniques are certainly among the most innovative techniques and with an incredible future potential supported by an ever more computational power availability. In this paper, various supervised learning techniques are investigated in order to quantify the generalized corrosion thickness that could occur in a steel structure. At this regard, a finite element model has been developed in the Finite Element Software OpenSees in which a stochastic ergodic white noise process at the column base has been introduced to simulate environmental vibrations. Furthermore, to simulate the operating conditions, the masses were introduced in probabilistic terms. Once the vibration signals were extracted in terms of accelerations, various features in frequency domain were calculated for AI training. Different configurations of Support Vector Machines (linear, quadratic, cubic and Gaussian) were explored also with Bayesian optimization of the hyperparameters and definition of the penalty matrix. The results obtained show the potential of these techniques for structural monitoring and their future possible extension to different structural typologies and with different damage types.