A deep learning approach to impact localization and uncertainty assessment in CFRP composites using sparse PZTs: Integrating experiments and simulations

Abstract

The propagation of elastic waves in CFRP composites is dispersive and multimodal, making impact localization using wave signals challenging. An end-to-end deep learning model with an encoder-decoder architecture was developed for impact localization in CFRP composites, using sparse piezoelectric ceramic transducer (PZT) arrays and a data-driven approach to enable online impact sensing. The model used a segmented training strategy and transfer learning to capture shared features between experimental and simulated data. According to the piezoelectric equation, it links experimental piezoelectric signals with simulated stress responses. The results show that the feature encoder trained can extract their shared features effectively. Meanwhile, the model successfully applied the laws from simulation data to the localization of experimental impacts based on the fine-tuning strategy, alleviating the challenge of limited experimental data. The prediction results in the test set show that good generalization, and impact localization can be achieved in milliseconds during inference. The average error in localization is only 3.42 mm over a 100 mm × 100 mm monitoring area. Compared to the traditional transfer strategy, the three-stage training proposed shows better generalization by constraining the encoded features. In addition, the Monte-Carlo dropout strategy is used to assess prediction uncertainty, analyzing the effects of dropout rate and repeated predictions on confidence intervals. The study provides a prospective solution for large CFRP structures to achieve fast localization of impacts under sparse PZT arrays.