Experimental and Computational Materials Defects Investigation

Production of railway axles (i.e., one of the basic material of the modern train) is an elaborate process unfree from faults and problems. Errors during the manufacturing or the plies' overlapping, in fact, can cause particular flaws in the resulting material, so compromising its same integrity. Within this framework, ultrasonic tests could be useful to characterize the presence of defect, depending on its dimensions. On the contrary, the requirement of a perfect state for used materials is unavoidable in order to assure both transport reliability and passenger safety. Therefore, a real-time approach able to recognize and classify the defect starting from the finite element simulated ultrasonic echoes could be very useful in industrial applications. The ill-posedness of the so defined process induces a regularization method. In this paper, a finite element and a heuristic approach are proposed. Particularly, the proposed method is based on the use of a Neural Network approach, the so called "learning by sample techniques", and on the use of Support Vector Machines in order to classify the kind of defect. Results assure good performances of the implemented approach, with very interesting applications.