Abstract:
Non-destructive ultrasonic inspection is employed in the railroad industry to inspect in loco the rods and couplings that connect wagons to each other in order to detect critical cracks in the composition during predictive maintenance. The test requires care because not only cracks are detected, but also natural discontinuities of each part. Also, a repair workshop may contain multiple inspectors, a factor that introduces subjectivity to the process. An Artificial Neural Network can aid in the identification of critical discontinuities, given training based on previous analyzes. The aim of this study was to elaborate a Python-programmed model capable of predicting, using images obtained by the ultrasound equipment, the presence or not of critical cracks in each report. A database containing intensity values for each report (input variables), and binary values related to the existence or not of critical cracks (output variable) was elaborated from processing 1222 images generated by the Phasor XS (General Electric) equipment. The data recorded were inserted into a multi-layer perceptron network, of 20-36-1 configuration and logistic activations. Partial results point to a 97% prediction accuracy, value that is compatible with similar studies, and support-metrics that validate model learning.