CNN 0.8 0.6 0.4 0.2 MLP FCNN RESNET INCEPTION ENCODER CNN 1.0 0.8 0.6 0.4 0.2 0.0 MLP FCNN RESNET INCEPTION ENCODER CNN 1.0 0.8 0.6 0.4 0.2 MLP FCNN RESNET INCEPTION ENCODER Figure 10. The distribution of: (a) accuracy (b) precision and (c) recall from a tenfold cross-validation for six classifiers on the impact test dataset. Models without transfer learning are indicated by red bars, while those with transfer learning are shown in blue. 0.8 0.6 0.4 0.2 0.0 CNN MLP FCNN RESNET INCEPTION ENCODER CNN 0.8 0.6 0.4 0.2 0.0 MLP FCNN RESNET INCEPTION ENCODER CNN 0.8 0.6 0.4 0.2 MLP FCNN RESNET INCEPTION ENCODER Figure 11. The distribution of: (a) accuracy (b) precision and (c) recall from a tenfold cross-validation for six classifiers on the PLB test dataset. Models without transfer learning are indicated by red bars, while those with transfer learning are shown in blue. ME |AI/ML 82 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 3 2307 ME July dup.indd 82 6/19/23 3:41 PM Accuracy Precision Recall Accuracy Precision Recall

showed minimal variation in performance precision was slightly higher without transfer learning, while recall remained unchanged. Accuracy was slightly improved with transfer learning. As for the PLB dataset, the CNN and MLP with transfer learning slightly outperformed their counterparts without transfer learning. FCNN underperformed with an accuracy of less than 0.1, while transfer learning further deteriorated its performance. Again, ResNet showed significant improve- ment through transfer learning. Unlike the Impact dataset, Inception with transfer learning showed slightly worse perfor- mance compared to without transfer learning. Encoder, similar to CNN and MLP, had slightly higher precision, recall, and accuracy with transfer learning. The observations from this study can be explained by the fundamental advantage of transfer learning, which can be explained by the reusability of the learned features. Models without transfer learning, though adept at discriminative patterns from training data, face difficulties in generalizing to unfamiliar data. This process often results in memorizing training data rather than assimilating generalizable patterns, thereby leading to elevated validation losses. On the contrary, models employing transfer learning derive initial benefits from patterns and features harvested from an extensive simulated dataset. These models exhibit reduced initial loss values, indi- cating that the simulated dataset provides a beneficial starting framework for interpreting the limited experimental data. Furthermore, fine-tuning allowed these models to adapt to the specific characteristics of the experimental data, resulting in significant improvement over epochs and better generalization capabilities. The distinct performance outcomes of different models, as illustrated by statistical metrics and visualizations, underscore the crucial role of model architecture in harnessing the effectiveness of transfer learning. Conclusions This paper proposes a novel data-driven approach to accu- rately localize two types of acoustic emission sources in an aluminum plate using six deep learning models: CNN, MLP, FCNN, Inception, ResNet, and Encoder. The models incorpo- rate deep transfer learning techniques to enhance their effec- tiveness in identifying the source of acoustic emission signals. The deep learning models were trained and evaluated using simulations of impact and PLB tests with a distributed sensor array designed to maximize information acquisition from the simulations. The results demonstrate the efficacy of deep neural networks with transfer learning in mapping acoustic emission waveforms to their sources and uncovering valuable insights from the simulations. However, this study’s limitation is the inability to identify the exact coordinates of the sources of the acoustic emissions. Future research should optimize the deep neural networks using larger training datasets and explore automated solutions like numerical simulations or robotic solutions to address this limitation. Additionally, while in this study Hsu-Nielsen tests were used to simulate fatigue cracks, further research should conduct more formal tests on actual propagating cracks to verify the performance of the proposed deep learning approaches under real states of stress. These efforts could lead to the development of more robust and accurate deep learning models for acoustic emission source localization in real-world applications. ACKNOWLEDGMENTS The authors would like to express their gratitude for the research support provided by the US Office of Naval Research under Award No. N00014-20- 1-2649 and technical guidance from Program Manager Dr. Ignacio Perez. REFERENCES Agarwal, N., A. Sondhi, K. Chopra, and G. Singh. 2021. “Transfer learning: Survey and classification.” Smart Innovations in Communication and Computational Sciences. Advances in Intelligent Systems and Computing, vol 1168. Springer, Singapore: 145–155. https://doi.org/10.1007/978-981-15- 5345-5_13. Ai, L., V. Soltangharaei, M. Bayat, B. Greer, and P. Ziehl. 2021. “Source localization on large-scale canisters for used nuclear fuel storage using optimal number of acoustic emission sensors.” Nuclear Engineering and Design 375. https://doi.org/10.1016/j.nucengdes.2021.111097. Bengio, Y. 2012. “Deep learning of representations for unsupervised and transfer learning.” JMLR: Workshop and Conference Proceedings 27:17–37. Chen, S.-X., L. Zhou, Y.-Q. Ni, and X.-Z. Liu. 2021. “An acoustic- homologous transfer learning approach for acoustic emission–based rail condition evaluation.” Structural Health Monitoring 20 (4): 2161–81. https:// doi.org/10.1177/1475921720976941. Cheng, L., H. Xin, R. M. Groves, and M. Veljkovic. 2021. “Acoustic emission source location using Lamb wave propagation simulation and artificial neural network for I-shaped steel girder.” Construction &Building Mate- rials 273. https://doi.org/10.1016/j.conbuildmat.2020.121706. Ciampa, F., and M. Meo. 2010. “Acoustic emission source localization and velocity determination of the fundamental mode A0 using wavelet analysis and a Newton-based optimization technique.” Smart Materials and Struc- tures 19 (4): 045027. https://doi.org/10.1088/0964-1726/19/4/045027. Cuadra, J., P. A. Vanniamparambil, D. Servansky, I. Bartoli, and A. Kontsos. 2015. “Acoustic emission source modeling using a data-driven approach.” Journal of Sound and Vibration 341:222–36. https://doi.org/10.1016/j. jsv.2014.12.021. Delashmit, W., and M. Manry. 2005. “Recent developments in multilayer perceptron neural networks.” Proceedings of the 7th Annual Memphis Area Engineering and Science Conference. Ebrahimkhanlou, A., and S. Salamone. 2017. “Probabilistic location esti- mation of acoustic emission sources in isotropic plates with one sensor.” Health Monitoring of Structural and Biological Systems 2017. 2017. https:// doi.org/10.1117/12.2258618. Ebrahimkhanlou, A., and S. Salamone. 2018. “Single-Sensor Acoustic Emission Source Localization in Plate-Like Structures Using Deep Learning.” Aerospace (Basel, Switzerland) 5 (2): 50. https://doi.org/10.3390/ aerospace5020050. Ebrahimkhanlou, A., B. Dubuc, and S. Salamone. 2019. “A generaliz- able deep learning framework for localizing and characterizing acoustic emission sources in riveted metallic panels.” Mechanical Systems and Signal Processing 130:248–72. https://doi.org/10.1016/j.ymssp.2019.04.050. Garrett, J. C., H. Mei, and V. Giurgiutiu. 2022. “An Artificial Intelligence Approach to Fatigue Crack Length Estimation from Acoustic Emission Waves in Thin Metallic Plates.” Applied Sciences (Basel, Switzerland) 12 (3): 1372. https://doi.org/10.3390/app12031372. Hamstad, M.A. 2007. “Acoustic emission signals generated by monopole (pencil-lead break) versus dipole sources: finite element modeling and experiments.” J. Acoustic Emission 25. J U L Y 2 0 2 3 • M A T E R I A L S E V A L U A T I O N 83 2307 ME July dup.indd 83 6/19/23 3:41 PM