Fatigue life prediction of overhead conductor wires using a physics-guided neural network

Previous works by the authors showed that Artificial Neural Networks (ANN) using as input parameters (i) the normal stresses considered in the SWT critical plane model, (ii) the material critical distance and (iii) the material yield stress were capable to accurately estimate lives for classical cylinder and sphere against plane fretting fatigue tests. The present study investigates the generalization capability of such a physics-guided ANN model when applied to a fundamentally different problem involving overhead conductor wires subjected to fretting fatigue. The model, originally trained using aeronautical aluminum alloys and classical contact geometries, is here evaluated under substantially different conditions, including wire-on-wire contact geometry, reduced length scales, different aluminum alloy families (AA1120, AA1350 and AA6201), and both constant and variable amplitude loading histories. Importantly, no retraining or parameter recalibration is performed. The results show that the proposed model provides consistent fatigue life predictions across all investigated datasets, with most results falling within a factor-of-three band and band mean error factors close to 2.0. These results are comparable to those obtained using a calibrated TCD-based SWT approach. The findings demonstrate that the combination of physically-informed inputs and nonlocal stress representation enables robust cross-domain generalization of the ANN model. This suggests that the trained ANN model can be applied for life estimation within a wide range of fretting configurations and materials.