ABSTR ACT This paper presents a novel data-driven approach to localize two types of acoustic emission sources in an aluminum plate, namely a Hsu-Nielsen source, which simulates a crack-like source, and steel ball impacts of varying diameters acting as the impact source. While deep neural networks have shown promise in previous studies, achieving high accuracy requires a large amount of training data, which may not always be feasible. To address this challenge, we investigated the applicability of transfer learning to address the issue of limited training data. Our approach involves transferring knowledge learned from numerical modeling to the experimental domain to localize nine different source locations. In the process, we evaluated six deep learning architectures using tenfold cross- validation and demonstrated the potential of transfer learning for efficient acoustic emission source localization, even with limited experimental data. This study contributes to the growing demand for running deep learning models with limited capacity and training time and highlights the promise of transfer learning methods such as fine-tuning pretrained models on large semi-related datasets. KEYWORDS: acoustic emission, deep neural network, finite element modeling, transfer learning, fiber optics, source localization Introduction Acoustic emission source localization is crucial in struc- tural health monitoring (SHM) and proactive maintenance of metallic structures. The constraints in deploying acoustic emission testing (AE) sensor arrays in real-world structures necessitate a shift toward intelligent, automated single-sensor approaches. Holford et al. (2001) pioneered the application of AE for damage location in steel bridges, establishing its impor- tance in SHM. Ebrahimkhanlou and Salamone (2017) further examined acoustic source localization and its significance in determining the origin of acoustic emission waves and assess- ing damage severity. Cheng et al. (2021) developed an acoustic emission source localization method using Lamb wave propa- gation simulation and artificial neural networks, proving effec- tive in I-shaped steel girder inspections. Ai et al. (2021) studied source localization on large-scale canisters used for nuclear fuel storage, addressing the need for optimal AE sensor deploy- ment. Ciampa and Meo (2010) proposed an approach using wavelet analysis and a Newton-based optimization technique for acoustic emission source localization and velocity determi- nation, contributing to the broader understanding of acoustic emission wave propagation and source detection. Significant progress has been achieved in acoustic emission source localization through the application of deep learning, demonstrating its promise in localizing acoustic emission signals (LeCun et al. 2015). Ebrahimkhanlou and Salamone (2018) proposed a deep learning approach for localizing acoustic emission sources using a single sensor in plate-like structures. This was further advanced by Ebrahimkhanlou et al. (2019), who introduced a deep learning–based framework for localizing and characterizing acoustic emission sources in metallic panels using only one sensor. Garrett et al. (2022) utilized artificial intelligence for estimating fatigue crack length from acoustic emission waves, a significant step forward in damage localization and quantification. Despite the challenge of false positives, the fusion of artificial intelligence and AE holds promising opportunities for enhancing SHM (Verstrynge et al. 2021 Hassan et al. 2021). A key challenge in using supervised learning algorithms for acoustic emission source localization is the difficulty in accessing labeled acoustic emission signals for existing struc- tures. Transfer learning is a strategy that assists the super- vised learning task when available training data is limited ACOUSTIC EMISSION SOURCE LOCALIZATION USING DEEP TRANSFER LEARNING AND FINITE ELEMENT MODELING– BASED KNOWLEDGE TRANSFER XUHUI HUANG*, OBAID ELSHAFIEY*, KARIM FARZIA†, LALITA UDPA*, MING HAN*, AND YIMING DENG*‡ *Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI † Nikon Inc., 9453 Innovation Dr., Manassas, VA ‡ Corresponding author: dengyimi@egr.msu.edu Materials Evaluation 81 (7): 71–84 https://doi.org/10.32548/2023.me-04348 ©2023 American Society for Nondestructive Testing NDTTECHPAPER |ME 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 71 2307 ME July dup.indd 71 6/19/23 3:41 PM

(Agarwal et al. 2021). Various studies have demonstrated the value of transfer learning in enhancing neural networks for acoustic emission source localization and SHM, such as Chen et al. (2021), who proposed an acoustic-homologous transfer learning approach for rail condition evaluation, and Hasan et al. (2019), who utilized transfer learning for reliable bearing fault diagnosis under variable speed conditions. Deep learning and transfer learning methods have shown great potential in improving acoustic emission source local- ization efficiency (Sun 2020 Bengio 2012). Ismail-Fawaz et al. (2022) presented a deep learning approach for time series classification using hand-crafted convolution filters, further enhancing AE capabilities. Ismail Fawaz et al. (2018) explored transfer learning for time series classification, while Zhang et al. (2017) studied the learnability of fully connected neural networks. Weiss et al. (2016) provided a survey of transfer learning, and Bengio (2012) emphasized the importance of deep learning representations for unsupervised and transfer learning. Though significant advancements have been made in applying deep learning and transfer learning to acoustic emission source localization, continued development and opti- mization of these methodologies are essential for addressing inherent challenges and maximizing their potential in SHM (Bengio 2012 Sun 2020). In this study, our principal innovation lies in the success- ful implementation of transfer learning through the pretrain- ing of six deep learning models on a large simulated acoustic emission dataset. This enabled the localization of acoustic emission sources using a single sensor. We pretrained con- volutional neural network (CNN), fully convolutional neural network (FCNN), Encoder, ResNet, Inception, and Multi-layer Perceptron (MLP) models using data from finite element method (FEM) simulations of acoustic emission impulses. Through transfer learning, we fine-tuned the pretrained models on the experimental dataset, improving their perfor- mance while reducing the number of experiments needed. Our results show that the pretrained models generalized well to variations in acoustic emission signals and could be applied to different model architectures and datasets. Overall, our research highlights the potential of deep learning techniques, particularly transfer learning, for improving the accuracy and efficiency of acoustic emission source localization. These findings can significantly benefit the development of reliable and cost-effective SHM strategies and are readily applicable to other nondestructive evaluation problems. This paper is organized into four main sections. The first section provides an overview of the laboratory experiments conducted utilizing pencil lead break (PLB) and impact tests at nine distinct positions. In the next section, to aid under- standing, data visualization is furnished through raw waveform plots of both simulated and real-life experimental data derived from impact and PLB testing. Additionally, a 2D t-SNE plot is provided to better illustrate the clustering structure of signals originating from nine distinct locations or classes. The third section introduces six distinct deep learning models, including our own, which were designed through the iterative fine-tuning of layers with unique training parameters. The architectural details of both the classifier and the transfer elements of our model are thoroughly analyzed in this section. The final section presents the results obtained by training these fine- tuned models using tenfold cross-validation. To give a compre- hensive view of the models’ performance, the mean loss and range of loss for each classifier, as well as for the impact and PLB tests, are plotted. The efficacy of each fine-tuned model is further evaluated by computing and representing key metrics such as precision, recall, and accuracy in a box plot format. Methods and Experiments The primary objective of the conducted experiments was to scrutinize the effectiveness of the suggested source localization techniques, utilizing a singular AE sensor, on an aluminum plate. As represented in Figure 1, the experimental setup com- prised a sensor, constituted by two frail fiber Bragg gratings (FBGs), forming a low-finesse Fabry-Perot interferometer (FPI) on a coiled single-mode fiber. This arrangement facilitated the detection of ultrasound on a solid surface. The setup employed a narrow-linewidth diode laser with wavelength tunability, designed to direct light toward the FBG-FPI sensor via a circu- lator developed in Karim et al. (2021). Before reaching the sensor, the light was passed through a three-paddle polarization controller, which facilitated manual adjustments to the laser polarization. The light reflected from the sensor was then directed to a photodetector (PD) through the same circulator. To obtain acoustic emission signals of higher quality, the output from the PD was amplified and filtered using a 50–500 kHz band-pass filter. Additionally, noise removal techniques, such as adaptive filtering, were employed to reduce any extraneous signals present during data collec- tion. It is selected based on its ability to effectively remove noise while preserving the signal of interest. The filtered and noise-free AE signals were subsequently utilized to train and test the deep learning models for source localization. Acoustic emission is a physical occurrence linked to stress waves, initiated by the abrupt liberation of elastic energy during the formation of cracks or damages within materials. AE signals can be captured and logged by attaching AE sensors to the sample surface. The AE monitoring process involves the collec- tion and analysis of these signals to assess the condition of the object under study. The Hsu-Nielsen PLB test, a widely accepted artificial method for acoustic emission signal generation (Sause 2011), was used in this study. It involves breaking pencil leads on a surface with an affixed AE sensor. For this study, PLB tests were conducted on a 2.54 mm thick aluminum plate measuring 0.30 × 0.30 m. The plate was partitioned into nine distinct loca- tions as delineated in Figure 2. Each of the nine representative points, denoted by a red dot, underwent the PLB test 10 times, using a 2H mechanical pencil with a 0.5 mm diameter lead. Furthermore, impact-like signals were gathered by dropping steel balls (4.7 mm diameter) from a height of 25 mm at the same AE sensor location illustrated in Figure 2. The ME |AI/ML 72 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 72 6/19/23 3:41 PM