Transfer learning approaches for fatigue life prediction of PVD-coated electro-conductive textiles with limited data

The reliability of flexible electro-conductive layers is a critical factor in the development of smart textiles (textronics). However, experimentally determining their fatigue life is a time-consuming and expensive process. This paper proposes the use of Long Short-Term Memory (LSTM) recurrent networks in a “many-to-one” configuration, utilizing Transfer Learning and Few-Shot adaptation techniques. The model was trained and tested on silver (Ag) and gold (Au) samples deposited via Physical Vapor Deposition (PVD) onto composite textile substrates (Cordura and Membrane), utilizing experimental data from prior research. To address the challenge of limited experimental data availability, we applied a Deep Learning framework capable of extrapolating non-linear resistance trajectories based on short measurement histories and signal derivatives (degradation velocity and acceleration). The analysis revealed that the type of metal used is the decisive factor for model transferability, while the mechanical properties of the textile substrate are of secondary importance. The results indicate that transfer learning can provide accurate resistance-trajectory forecasts for selected source–target configurations, particularly when their degradation dynamics are sufficiently similar. However, the method performed poorly for Membrane-Au, demonstrating that severe degradation-timescale mismatch can prevent successful transfer. The proposed methodology can enable effective life-cycle estimation of textronic materials at the early stage of operation, significantly reducing the need for prolonged destructive testing.