MACHINE LEARNING TECHNIQUES FOR ACOUSTIC DATA PROCESSING IN ADDITIVE MANUFACTURING IN SITU PROCESS MONITORING A REVIEW HOSSEIN TAHERI* AND SUHAIB ZAFAR† ABSTR ACT There have been numerous efforts in the metrology, manufacturing, and nondestructive evaluation communities to investigate various methods for effective in situ monitoring of additive manufacturing processes. Researchers have investigated the use of a variety of techniques and sensors and found that each has its own unique capabilities as well as limitations. Among all measurement techniques, acoustic-based in situ measurements of additive manufacturing processes provide remarkable data and advantages for process and part quality assessment. Acoustic signals contain crucial information about the manufacturing processes and fabricated components with a sufficient sampling rate. Like any other measurement technique, acoustic- based methods have specific challenges regarding applications and data interpretation. The enormous size and complexity of the data structure are significant challenges when dealing with acoustic data for in situ process monitoring. To address this issue, researchers have explored and investigated various data and signal processing techniques empowered by artificial intelligence and machine learning methods to extract practical information from acoustic signals. This paper aims to survey recent and innovative machine learning techniques and approaches for acoustic data processing in additive manufacturing in situ monitoring. KEYWORDS: additive manufacturing, in situ monitoring, acoustic, machine learning, data processing Introduction Various additive manufacturing (AM) methods are utilized for manufacturing parts with complex geometries and compli- cated features that are either unfeasible or highly challenging to produce via traditional manufacturing techniques. This outstanding capability of AM provides substantial design flex- ibility and facilitates the production of complex parts with marginal added cost compared to subtractive and traditional manufacturing methods (Calta et al. 2018). Laser powder bed fusion (LPBF), directed energy deposition (DED), and wire arc additive manufacturing (WAAM) are among the most popular methods of metal AM (Koester et al. 2018). Fused deposi- tion modeling (FDM), stereolithography (SLA), direct ink writing (DIW), and selective laser sintering (SLS) are the most common AM techniques for polymers (Baechle-Clayton et al. 2022 Lee et al. 2020). The AM processes not only can cause different mechani- cal properties for the parts manufactured, but also lead to the potential generation of specific types of discontinuities and defects in AM parts (Koester et al. 2018, 2019b Taheri et al. 2017). The types of defects in AM parts significantly depend on manu- facturing process conditions and type of materials. A summary of defect types, causes of defect generation, and their potential effect on AM parts is presented in Table 1. Although inspection and quality assessment for the manu- factured parts can be done after the production is finished (ex situ), there are several significant challenges in traditional ex situ inspection methods. One of the major challenges of tra- ditional inspection of AM parts is due to the capability of AM techniques to produce complex-geometry components. This is an outstanding capability for AM but makes traditional inspec- tion of AM parts extremely challenging since many available nondestructive testing (NDT) techniques have been developed for simpler geometries (Bond et al. 2019). Another primary concern in post-production or ex situ inspection of AM parts is that AM techniques are used to manufacture many critical, high-valued, or exotic parts. Possible rejection of such unique parts due to unacceptable quality causes a significant loss of time and cost and is not a desirable outcome for industries (Koester et al. 2018c Taheri 2018). Despite the complexity of the processes in AM, the layer-by-layer deposition of materials allows the measurement and recording of large amounts of data on each layer for statistical process monitoring and quality assessment (Grasso and Colosimo 2017 Koester et al. 2018b). *The Laboratory for Advanced Non-Destructive Testing, In-situ Monitoring and Evaluation (LANDTIE), Department of Manufacturing Engineering, Georgia Southern University, Statesboro, GA, USA 30458 htaheri@georgiasouthern.edu † Stellantis, Chrysler Technology Center, 800 Chrysler Dr., Auburn Hills, MI 48326, USA Materials Evaluation 81 (7): 50–60 https://doi.org/10.32548/2023.me-04356 ©2023 American Society for Nondestructive Testing ME |REVIEWPAPER 50 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 50 6/19/23 3:41 PM

In situ measurement and monitoring techniques using various sensors and NDT methods have been extensively utilized and studied over the last few years for understanding and predicting the alterations in AM process parameters and, consequently, the quality of the manufactured parts. In situ measurement data obtained over the entire period of manufac- turing processes, combined with ex situ material characteriza- tion and information from process modeling and simulation, are essential for reducing the time and cost of process develop- ment, improving part quality, and minimizing defect formation (Hossain et al. 2022 Koester et al. 2018b). A large body of existing and rapidly growing literature is devoted to in situ monitoring methods, surveying various in situ monitoring techniques and sensors used for different types of AM processes. High-speed visible imaging (Scipioni Bertoli et al. 2017), thermography (Raplee et al. 2017), and X-ray imaging (Calta et al. 2018) are among the most used methods for in situ process monitoring for AM. Optical-based in situ monitoring methods can monitor process conditions and variations on the surface of the parts but are limited in assessing bulk material behavior. In addition, high-resolution imaging at high scanning velocities requires an external illu- mination source (Lott et al. 2011). Also, a wide range of mag- nification may be needed to cover the imaging of the entire melting pool (Lott et al. 2011). Arntz et al. (2018) analyzed the melt flow dynamics of a laser cutting process by in situ high-speed video diagnostics (100 000 fps). They showed a correlation between fluid dynamics, cutting velocity, and the average roughness of the cut flank (Arntz et al. 2018). In contrast, X-ray-based measurement methods can penetrate the materials and provide valuable information regarding the structure of the part. However, the complexity and cost of the X-ray monitoring technique and availability to most industries and manufacturers for widespread implementation of AM is a significant challenge. On the other hand, acous- tic-based techniques have been used historically for a variety of process monitoring and part qualification applications, such as in the welding process, where its rapid solidification phenomena are very similar to the AM process (Taheri 2018). Recent work has investigated the potential application of acoustic emission testing (AET) for AM processes (Koester et al. 2016, 2018a, 2019a). Accordingly, despite the type of sensing and measurement technique used for in situ AM process monitoring, analyzing the recorded dataset to identify, map, and potentially charac- terize the defects will be the next challenging step. The large dataset size and real-time processing are significant challenges in processing data for in situ measurement (Taherkhani et al. 2022). Artificial intelligence (AI) and machine learning (ML) algorithms can be promising solutions for such problems (Taheri et al. 2022). Researchers have used various supervised (Gobert et al. 2018), unsupervised (Scime and Beuth 2018), and reinforcement learning algorithms (Knaak et al. 2021) for the prediction of defects during AM processes. AI/ML methods have significant potential to improve the AM processes and the quality of manufactured parts. The T A B L E 1 A summary of common process-induced defects, their causes, and potential effects on part quality in laser metal additive manufacturing (Herzog et al. 2023) Defect type Common causes Potential effects Keyhole pores • Excessive input energy density • Reduction in mechanical properties • Reduction in fatigue properties Lack of fusion pores • Insufficient input energy density • Reduction in mechanical properties • Reduction in fatigue properties Gas pores • Gas entrapped in feedstock • Gas entrained into the melt pool • Reduction in fatigue properties Cracking and delamination • Residual stresses exceeding the local ultimate tensile strength • Insufficient bonding between layers • Part failure Deformation • Residual stresses exceeding the local yield stress • Conformance failure Alloy compositional variance • Improper powder deposition • Differing chemical mobility • Preferential evaporation • Gas incorporation/adsorption • Inhomogeneous mechanical properties Balling • Low/high input energy density • Surface oxidation • Part/conformance failure • Formation of other defects Rippling • Instabilities of layer-to-layer deposition • Part failure • Production failure Spatter/particle ejection • Overheated melt pool • Recoil pressure and melt plume • Formation of other defects 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 51 2307 ME July dup.indd 51 6/19/23 3:41 PM