Fatigue degradation of metallic structures is caused by repeated load cycles. In major part of the lifetime, the stresses and strains cause in local microstructural changes, and only in the last minor part, macroscopic cracks are formed and lead to a final failure. However, established in-service inspections rely on techniques that are qualified to detect larger cracks. Therefore, the identification of the fatigue-induced microstructural changes by specific sensors has the perspective of an effective early fatigue monitoring.

In this contribution, the fatigue behaviour of the austenitic stainless steel AISI 347 is investigated under different temperature and medium conditions. The microstructural changes including phase transformation from austenite to α’-martensite, can be characterized by the change of electrical and magnetic properties. Changes in electrical resistance and magnetic field strength can be measured by dc potential and eddy current probes. These measurement techniques are used in specimen-level fatigue tests and the measurement signals are recorded over the fatigue lifetime.

While the sensors provide detailed data allowing the extraction of features for machine learning techniques, the identification of material scientific informed labels is a challenge, since the direct characterization of the microstructural state is not possible as an online monitoring technique. Therefore, the fatigue life in the test itself and the implied linear damage accumulation parameter is chosen as a label, while the sensor signal evolution is interpreted qualitatively regarding the microstructural phenomena.

With the extracted features and labels, a regression analysis is applied based on an artificial neural network, using TensorFlow. The predictive capabilities are shown, and the application of the trained model for other test types is investigated.