Evaluating Transfer Learning Strategies for Neural Network-based Impact Location Model
Daniel del-Río-Velilla, Jesús Sesé, Fernando Sánchez Iglesias, Antonio Fernández López
TL;DR: This paper investigates transfer learning (TL) as a strategy to adapt a multilayer perceptron (MLP) trained on a stiffened AS4/8552 CFRP panel to ten alternative sensor layouts, simulated via controlled sensor-index permutations grouped into three families of increasing severity.
Passive impact localization using piezoelectric sensor (PZT) networks and machine learning is an established approach for structural health monitoring of composite aerospace panels. A key practical limitation is that models trained on one sensor layout fail when deployed on a structurally similar panel with a different instrumentation arrangement, and full retraining is costly. This paper investigates transfer learning (TL) as a strategy to adapt a multilayer perceptron (MLP) trained on a stiffened AS4/8552 CFRP panel to ten alternative sensor layouts, simulated via controlled sensor-index permutations grouped into three families of increasing severity: intra-column swaps, intra-row swaps, and full-plate reflections. Six TL strategies, from output-head-only retraining to full fine-tuning, are benchmarked at five adaptation data fractions against a from-scratch baseline. Results show that full fine-tuning is the only strategy that consistently recovers baseline localization accuracy (25.7~mm mean Euclidean error) across all permutation families, requiring as little as 5\% of the adaptation dataset for mild and moderate shifts and 25\% for severe full-plate reflections. Partial fine-tuning strategies fail to recover acceptable accuracy at any data fraction, due to the globally distributed nature of sensor-layout encoding in hand-feature MLPs. These findings provide practical guidelines for redeploying impact localization models across panels with varying sensor configurations.