Scribd Logo
Upload

Welcome to Scribd!

  • LPBF A Review On The Four Drivers For An Online Control (2023)

    Uploaded by

    Reginald Elvis
    15 views17 pages

    Original Title

    LPBF a review on the four drivers for an online control (2023)

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    15 views17 pages

    LPBF A Review On The Four Drivers For An Online Control (2023)

    Uploaded by

    Reginald Elvis

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    of 17
    Journal of Manufetaring Processes 103 (2028) 412-429 ‘Contents lists lable at ScienceDirect Journal of Manufacturing Processes journal homepage: yw: eleevier com/loeatelmannro ELSEVIER Review y Laser powder bed additive manufacturing: A review on the four drivers for an online control Francesco Lupi"’, Alessio Pacini”, Michele Lanzetta” bermenof frat ingens, 56122 es, ety * parent of Chil ant na ner, S512 Pa, aly Keone Online conta of Aditive Manufacturing (AM) processes appear © Be che next challenge inthe ansion iano cent toward Industry 40 (140), Altnovgh many efforts have been cedcnted by industry and research in the ar ‘ecndes, tere remain bean room for improvement, Aditionally, the exiting sient Ueatre lacks 9 Power Bed Psion (PHP), wit particular emphasis on two sobeategoie, namely Selective Laser Sintering (G15) and Ssecive Later Mling (SLM). Through a systematic erature review, hi ate inl dened ‘double checked on Scholar. The remit were refined through mlple phases of incaslon/exeaon exter, review of four identified drivers) Online eontolable inp parameters) Online observable tpt nates, 58) Online sensing teenies i) Online feedback srtees, adopted frm the general Deming cote! loop PlanDo.CheeioAct (POCA). Ultimately, this aril delves ino the challenges and proses inherent ia he 1, Introduction Born almost forty years ago, Additive Manufacturing (AM) emerges sone ofthe most promising non-conventional manufacturing processes ang has since become a key enabling technology in the current Industry 4.0 (4.0) revolution [),2] AM allows for a higher degree of freedom in shaping, compared to traditional manufacturing methods, but it does ‘come with some trade-off, such as limited mechanieal properties and process proictivity (9), To date the aerospace, defense, and biomed ‘cal industries have made the most significant investments in this sector, diven by the demand for highly customized parts, followed by other sectors like tooling, jewelry, and automotive]. Around 2018, the hype surrounding AM subsided inthe mass media, but interest in esearch and ‘commercial applications has never been higher [5,0]. Notably, thou sands of companies are now leveraging AM, leading to 8 remarkable Industry expansion of 7.5 9 that resulted in nearly $12.8 billion in revenue in 2020 [7,8]. Within this dynamie landscape, this atile fo ‘eases on one specific class of AM technologies, namely Powder Bed Psion (PBF) [9], a depicted by the dotted box in Fie. 1 mal eden isco opp nip CF Lap tps org/10.1016/}mapr.2023.08.022 PBF was among the earliest and has remained one of the most ver satile AM processes, catering to polymers and metalsas well as ceramics, composite, and biomaterials [10]. Among several thermal sources, laser, and electron beams, respectively referred to as Laser Powder Bed Fusion (L-PBF) end Flectron Beam Powder Bed Fusion (E-PBF), are the most historically acknowledged techniques. The third typeof thermal source which has reeently entered the market uses Infra-Red (1) lamps [il]. Table 1 provides brief overview ofthe four subclasses of PBF, ‘categorized based on the thermal source, as highlighted in ie. 1 The focus of thisarticle, indicated by the blue/shaded labels in ig. 1 and Table 1, lies on L-PBF Gc, SIS and SLM), with particular emphasis ‘on metal powders due to their increasing significance (12,13. Online process contol for metal PBF is stil in its development stages [1]. Despite the sophistication of AM machines and digital cwin Capabilities, the process is often operated as an open loop, with users manually tuning process paramerers based on postprocess character leation and analysis [15,16]. While some authors and vendors have proposed actual online contol strategies, they ace limited, primarily relying on simple alarms or process blocking upon defect detection Received 1 August 2022; Reclvd in revises form 31 July 2028; Accepted 6 August 2023 1526-6125/0 2028 The Authors. Published by Fsvier Lid on behalf of The Society of Manufactring Enginers. This i an open access avec under the CC BY Aiente Cp /ereatvecommonsorlicenes/by/407. F hpteeat our of Manafcurg Frcs 108 (2023) 15-429 AM Technologies Binder Directed meray Material Materat onderBad Sheet a Jeting Deposition Seting Extrusion Fusion (PBF) Lamination merization i | rar] = oo weal ie pear wen SESE Ghaie FSS GA wow | SUHTe pe Ppl Fig. 1. The AM tecnololes classes, desrpons, subclasses, and pinary materi fom [tn Bue/shaded the two subelssesin-scopeof the present ate. For ath subslas, the extended names po ‘tom left orght Blader dering (0), Laser Eagncering Net Shape (LENS), Hleevoa Seam Addlve Manufacturing (EAM, noel Jeng (AL), Nano Pariee Jeting (NP), rep On Demand (DOD), Fused Depstila Medallag (FDMD, Mult Jet Fusion (MLF), Selective Laser Sintering (SL, Selective Laser Meling (SLM), Elecuon Beam Making (EBMD, Laminated Object Manulaetusing (LOM), Streolinogeaphy (SLA), Digital Light Processing (DLP), Continuous Dial Light Procesing (CDLP). (For interpretation ofthe references to calor in hs Figure legend, the reader s efered tothe web ‘asin of chs artle, (1718). This ariel aims 0 investigate current gaps of online control by reviewing the tate ofthe at of four major process drivers inspired from the Deming cycle [1°] loop conteol (Fi, 2) and following there View protocol proposed by [20 ‘As depicted by fig 2, online controllable input parameters are the starting point. These parameters are planned (.., P-Plan) in advance an then actuated during the production process (Le, D-Do), thereby ‘generating process signatures. The next ep involves monitoring these ‘online observable output signatures using appropriate insta sensors ‘., CCheck), Subsequently, the sensors provide a stream of data used to ‘adjust the input parameters in accordance with the Feedback strategies (he, A-Aed. ‘Our artiele builds wpon and extends the reviews conducted by (4,17,21-24], adopting aligned standard terminology, considering more recent papers, and following the structured approach in reviewing s¢- centile articles for each of the four drivers presented in Fig. 2. Our ‘overarching objective is to provide readers with a comprehensive and clear understanding ofthe complexities surrounding online control of metal L-PBF processes, tackling interdisciplinarty topics from a holistic perspective. The remainder ofthe article is organized as follows and is visually presented in Fig. 3. The scope and Research Questions (RQs) have been ‘defined as input fr the review process The subsequent article sections hhave been planned and implemented to ensure a logical and coherent flow (Fig. 3, center), andthe related contributions have been presented to address the inital RQS within the scope ofthis article (i. 3, right side) ‘While Selon 1 aims identifying the fest contebution of this atcle (ee, adopting the Deming cycle eateully customized to formalize the cnline control of L-PBE processes), Section? introduces the metal L-PBF process, providing the context ofthis work. Section ® details the strc tured procedure adopted to carry out the literature review on the four drivers, answering the frst RQU: what aspects do we aim to examine fn he lteraure? Section 4 reports and summarizes the state of the art on the four driver, ased onthe final st of 95 selected papers filtered from an Initial list of over 200 Scopus indexed papers. This section offers = Aetaled account ofthe literature, addressing RQ2: what are the primary ‘approaches found inthe eran? Ia Section 5, we provide a concise ‘overview and takeavays from the main aspects ofthe four drivers, along, ‘ith their eurent level of maturity, to answer ROB: what are the Key ndings from the lterure? Soon 6 focuses onthe future directions and ‘challenges conceming data collection, management, and processing for ‘etal L-PBF online control, addressing RO4: what are the main fare challenges? Although the focus of this article is on metal L-PBF, the readers can readily apply the methodology used inthis article, inckuding the four drivers, to other specific AM techniques and various materials. 2. Metal L-PBF process overview Inthis section, we presenta comprehensive overview of metal -PBE to provide context forthe current study. Sis. 4 summaries the general, ‘AM process (depieted in white), with the addition of an online closed Table The fou PRE processes, shor eseripson, thermal source and required power. In blue the proces inscope ofthe ar Sar eepon Thema Regeed em The EB proces dea with high power enery sal to SLM, butitusecleceon bas end ole aserbeselapeoach Eaceen gh my ‘ Fig. 2. Graphical one af the ale. The four drivers ae highliahted in he ‘boxes Each box has ben labeled vate ins of Demin cone lop tes (Plan, D-Do, Check, AAC) 1° control loop (shown in green) and Statistical Process Control (SPC)/ ‘monitoring (shown in orange). The highdevel ow diagram illustrates the different phases ofthe process (represented by solid boxes) and the ‘outputs (represented by doted boxes). Solid arrows indicate the logical flow ofthe proces, while dotted arzows represent the information iow. CControl-limit detection in SPC are widely recognized asthe industry standard for in-situ detection and control of L-PBF processes (25) ‘Traditional SPC control charts, pioneered by Shewhart, enable the our of Manafcurg Frcs 108 (2023) 15-429 establishment and maintenance of statistical control over critical out puts in complex manufacturing and process environments. By analyzing, samples and assessing quality characterstis, control charts help Iden tify changes in process levels and detect non-random patterns that indieate the need for intervention. An out-of-control sitsation occurs ‘when special causes, n addition to common-cause variations inherent in the process, begin to influence the process output. Control limits (Le, upper, and lower), calculated using different methods, serve as thresh ‘olds for sample data and prompt necessary actions such as process adjustment or investigation when measures fall outside these limits. Despite the potential demonstrated by using process limits as a viable and practical method for detecting undesirable conditions the focus of contol chars is on idensifying abnormal process conditions rather than ldentiying individual defects or implementing realtime process cor rections. Therefore, the alms provided by the recorded data are not sulicent to manage the inevitable and unpredictable variations ofthe process conditions and to ensure parts consistency. Additionally, these detection limits are often hardcoded and derived from experimental processes, operator experience, or simulation, rather than being learned from data, as isthe case with nev trending Machine Learning (ML) techniques [25]. For this reason, an advanced control strategy ig. 4 green) with closed online contol loop (red enclosed area) is, ‘needed [27]-In the following Scvions 2.1 and 22a brief description of ‘offline contto! loops and general AM activities is reported. This into duction alms to provide a comprehensive understanding of the ‘manufacturing process behind the four drivers reviewed for closed on line contol loop. 2.1. Requiremens, constrains and knowledge management During che part design stage (ig. 4, white, lefsie), a virwal 3 Dimensional (3D) model of the geometry intended for manufacturing is ‘reated using Computer-Aided-Design (CAD) software [28] or derived ‘rom sean data of an existing physical objec (2°) Subsequently, the 3D model is converted into a Standard ‘Tessellation Language (ST) file format, which describes the surface of the object using triangular facets Article inputs Article sections Article outputs fel Seton + Mensfeatin of main 4 divers for Klas online meta -PBF contol based on Scope nd background Deming yee input paraeters, sefinton r Sega] copa Sears semi edo) fi Metal t-PBE E+ ntoductory summary of met L- [Eeessaienerier DF process Sein} ROI: whararpects do we aim |__| Scope and to examine nthe literature? Ler roar Tmelod [| Review ffiteraure onthe & arch — (hans _, diver analysing 9 al papers Contrib auestions ROE wharan Weprimay |] | geewngsednasemng |” steve rom 217 Scopus approaches fund nthe 7 Besioa indexed wore tiers? he fourdivers | ROS: wha are te key [Scans Rtas 1], Summary ofthe ret on he 4 Siig fom the erate? fad dicasion | divers fr online contol view Seen RQ what arth man 1]. cursatebtenges and opportuni ich Outten }]—> as ec Fig 3. The atl stracsre according tothe scope and AQ inp article sections ofthe at of etl LPR. highligh in blue and ariel contibations as tp al fesse onthe state F npieeat our of Manafcurg Frcs 108 (2023) 15-429 Requirement [Bestel Koowedge JP moiies “+ and expense |Littniion || Stoeger |b cura QQ ‘Part design oe — wows 5 Fete neni |) Spates | Baan 4 ieee] Se . message or car = Cottle 4 re peacin | [m1 Cour rman” [pce = sash y aefa i 1 I eSPC/ Monitoring | Fig. 4 AM proces ow diagram with general activites in whit. Closed online conto lop jn green and industrial standards (with SPO) in orange. The grea red ‘nese aren represents the core ofthe eurrent atl. Offline data-driven contol modeling reresent he refinement an tining oop ofthe control, inlving ‘the development of relationships between contelable parameters ae observable signatures through adatriven approach, The orange area represents the coe of the SPC and monitering proces Olin char definition and refinement represents theafline ting fr SP, sneladng tks cha iit denon and refinement ‘The ofine lop is adopted for the modeling ofthe proces through a dtndrven approach, focsing on developing relationship bere controllable parameters and requirements centri (Fr intrpeation ofthe references t color thi igre, the rede i efecto the web version of this atl.) [80]. This mathematical separation of the 3D volumetric model inco slices generates @ machine-readable G-Code that contains all in struetions for the motors and other machine components, serving 88 additional input parameters (31,22) ‘As shown in Fig. 4 (whit, right-side, the production process re quires setting up contollable input parameters for the AM machine. ‘These parameters can be categorized into two main groups: ) predefined parameters that remain constant throughout the entire process and ‘online controllable input parameters, Predefined input parameters Include: laser type (€ OOz, Nes YAG, laser ber) [33-95], laser ‘operating mode (Le. continuous, pulsed) [6], laser beam quality factor [O71 powder properties (eg, material, size, thermophysical and flow. ability properties (98,091); inert gas type Ce. argon, nitrogen) [0]; and recoating mechanism (eg, sti seraper, sft squeegee, coer [1]. Detailed information about online controllable input parameters (.., ‘one of the four drivers reviewed in this article) will be provided in Section 4. Through proper knowledge management and offline loops (ig. 4, external loops), experimental correlation of input parameters and output signatures can be derived. These correlations ae utilized in ‘oth SPC and closed control approaches to achieve desired properties of ‘the manufeetured part, typically through experimeatal data-driven approach (eg. Design of Experiments DOE) []. SPC act asa sale guard by halting production if process signatures deviate beyond Acceptable limits. On the other hand, the closed control approach goes beyond mere monitoring when an issue arse. It involves a feedback loop that dynamically mocifies specifi controllable input parameters daring the manufacturing proces (42). This adjustment is done in real time to align the utp signatures with the desired ones. ‘The external oops can be associated also to the “long ran’ or “ys temic” corrective actions promoted by’ 1SO 9001 continuous improve ‘ment approach, which relies on developing process knowledge through objective evidence (Le, process logs) [4]. However, the development ofthese ofline cycles is beyond the scope ofthis article, and the desired output signatures as well as related input parameters are assumed known inthe following sections 2.2, Production process, postprocessing and quality control ‘As shoven in Pig 5, the production process is performed using L-PBE system (ie, AM machine) made by several sub-systems, which are Introduced inthe following. ‘The primary process subsystems include the optical chain, identified by the elements ©-O, and the powder feeding system, represented by clements @-®. In F's. 5, the (blue) arrows indicate the degrees of ‘reedom ofthe moving parts. The production process begins by loading the powder into the supply chambers @. The powder feed pistons and Daseplates © are initially positioned at the minimum excursion (Le, lowest position), hile the building piston and baseplate @ are set at their maximum excursion (Le, highest postion). Layer by layer, alte. nating left to right supply chambers &, the recoating mechanism ® spreads the powder feed supply @ on the top of the powder bed ®, ‘hich has been lowered bythe height of one layer moving @ down. A Fresh layer af powder coated on top ofthe previous layer is therefore scanned and locally Fused by the energy input ofa laser beam ©. The F npieeat our of Manafcurg Frcs 108 (2023) 15-429 ig. 5. The PBF system and is mala component. ©—laser source, Q—laser bea taser, —collmatr, ©—beam expander, @—scanning miter, @— Caheta leas, @—taser beams, O—powder bed, ©—recoting mechanism, @—buld pst and baseplate, @—powder fed piston and baseplate, @—powdes Teed supply, {@—Dullding chamber, @—powder suply camber laser beam i provided by laser source @, transfered bythe laser beam ‘wansfer @ to the opties! components such as collimator @ and beam ‘expander @, and deflected by a scanner system. The scanner system ilies xy moving micors © driven by galvanometers to accurately position the laser spot along the scanning path To ensure high precision, the scanner is liquid-cooled duet its sensitivity to thermal deformation. Finally, ftheta lens @ are usd to focus the laser beam. This sequence is repeated until the 3D object is completed inside the building chamber ©. Typically, the baseplate and powder bed are preheated, and the ‘building chamber © is protected with inert gas, Excess material can be disposed of in an overflow tank and reused in the next building job [45]. ‘fer the actual production process has been completed, several downstream steps are required to fnish the AM proces. Asilustrated in Fg. 4 fr Instance, the part must be post processed (eg, removal fom the baseplate and clesning of powder residues). Depending on the application of the manufactured past and the specific requirements, adeitional post processing steps such as machining, surface, heat, oF ‘ceemical reatment may also be necessary [45]. Finally, offline quality ‘control is caried out to assess the conformity of internal and external Features ofthe produced part (45,7) 3. Material and method ‘This research was conducted through a systema literature review, re.adapting the methodology proposed by [20] and widely use by other academics [48]. Inthe cucrent article, we introduced a novel deductive approach derived from the Deming cycle [291, which inspired us to ‘explore and categorize the four main drivers considered for online control of metal AM. Table 2 detals the steps followed fo this review and answers to ROL: what aspects do we aim to examine inthe erature? ‘Using the proposed esearch string, we performed a query on Scopus [ooh the largest datanese of peer-reviewed scientific literature [50]. A query isa structured sequence of words in the field of Information Retrieval (R) used to formalize the search of information (eg, scientific ‘papers, patents, websites) on a given database. Selected Keywords are connected by logical elements, such as “OR’, “AND, “NOT” [51]- After Phase 1-3 we intated the defined query on Sopus advanced search portal, yielding 217 contributions. Subsequently, Phase 4 involved refining. these contributions through automatic inclusion/exclusion ceiteria, resulting in @ reduetion to 139 papers. Inially, conference papers were excluded due to their typeally lower scientific impact and robustness. In Phase 5,2 thorough examination of tiles, abstracts, and ‘keywords led to a further reduction to 86 contributions. In Phase 6, we condlicted manual refinement via Google Scholar to triangulate the re sults, incorporating two relevant additional review papers. Recognizing, the advantages and limitations of conference papers, i, faster publ: cation versus reduced robustness, a threshold of 30 citations was ‘established for inclusion, Asa result, 7 conference papers were added. "The fnal set of 98 papers underwent an in-depth analysis to extract, {information pertaining to each of the four drivers (Le, input parameters, ‘output signatures, sensing techniques, feedback strategies) as per RQ2: what are the primary approaches fund in the erature? 4, The four drivers In this section, we classified the collected 95 papers into the four identified drivers according to Fig. 2. Table provides details about the Classified papers andthe corresponding artile sections that offer an in dep review of each driver. 4.1. Input parameters and output signatures overview Inthe metal LPBF process, many different parameters (i.e, over 50) impact the ultimate quality of the finished 3D part [22]. Such complesty creates a significant challenge in understanding process physi and developing an effective process contol strategy. According to [17], L-PBF process variables can be identified ax process parameters and process signatures F npieeat ‘Table? our of Manafcurg Frcs 108 (2023) 15-429 Review protocol according to 10}. Used databases: Seops 9] (and Googe Scholar fr wiangulation). 1 Reywors Menitoion 2. Quy sing deeopnent ‘he dees popes 5 Mana tant (exo) Gestion) The keyrords ene was driven by the soc prpow ade [pal proce, ad mar) hve ben Goad opti a he thal concepts tha mst be (al ened in he ue) eee The "OR- operon wae ano onan mantel sar ‘keywords fora specie cone nena). Pants ht despite ean and ne concep ar ogo, for abode quay weer ‘hes fw etre wa “OR” peat. Tie shar keyword cite The kent wine pepe ul mm hi eae were seed ae Scop et, {Caret heywords were ound oe bly competent ‘ren rer In pata the lowing icon condone were Sle Source ype Jarl OR Review OR one ree OF ‘ook capo Language np Time nde = 2007 Tay am) using hs phase he thes, aba nd keyword fhe 138 popes ‘rere aa emine to determine flere oo Fach ‘Spied objectives Incas of ueersny rpg the en tre cotent mo inspec ber main tbe deen a eove te Je hs phase, the same cary war acted in Goole Seblars ‘open the ete rough analtion A ol of dnl ‘eve ppc were hoen an inde i teil Ftherme, east 30 cops cations mere eviwed adage he slertorol7 oe ‘entree pp indie casas ech oe ow Oe: Toxo ana lng ins ab ‘min onne: elis on ine else ale) Goatees cont otk) Preece ser meting oe ser Sater pom ed (sta OF i sta" OR “tine” 08s” OR“ ine) OR (ontne OF eine” OR one” OR rele” OR ea ine") [AND Cantor” OR anal O8 fda) AND Cale se ‘lng’ OW select ner sintering” OR “power be aan” OR “Sis” 08 "P8F" OR SLSANDC eta) {TEMAbAaS-KevCinat OR i OR stn” OR sine” OR a ‘ie AND Crake” ON sone OR eedback) AND slate Iotermeting" OR oie nr string" OR poder be ao O8 "LM OR PBF OR'SLS7) AND Ce) (utp a pee lnpus papers ‘utp 95 pope fear avers Sess edo te. Process parameters are input 10 the process. They are either potentially conrollabl (e.g. laser power and spot size, seanning speed, layer thickness, scanning strategy) or predefined (eg, laser wavelength, powder bed material and type, inert gas type). The focus ofthis article will concern controllable parameters used to control the heating, ‘melting, and solidification proces, thus maintaining part quality during the process [171 Process signatures occur during the actual built and are often called “the voice of the process”. They are dynamic characteristics of the powder heating, melting, and solidification processes These signatures Aare categorized into two classes: observable signatures and derived sig ‘atures (17). The former ean be observed and measured during the process by using in-situ sensing devices and are considered in this, article. Controllable and predefined input parameters are related to the observable and derived output signatures (17). Thus, developing cor relations between input parameters and output signatures support actual closed-loop control, withthe goal of embedding process know! ‘edge into furure control approaches [1,17,21-23). This aspects into duced in Fig. 4 and detailed in Fig. 6, which recall the intial is. 2 by offering an overview ofthe two drivers. ‘As depicted in Fig. 5, the process online controllable input ps rameters (detailed in Sects 41-1) and eaine observable output sig natures (deseribed in Secions 4.1.2) were both arranged in three main subcategories (Le, laser, powder bed, and building ambient). For those readers interested in a broader overview on predefined parameters (in addition tothe ones highlighted in Section 2) and related signatures, ‘hich are notin the scope of this article, the following extensive reviews are suggested [17,221 In he following Table 4, the retrieved papers ae classified based on ‘the highlighted categories in Fig. 6, following the preliminary casi cation adopted in Table 5 for these wo érivers 44.1. Online controllable input parameters 441.1, Laser, As for online controllable laser parameters, the laser power (P) and the scanning speed (v) identified in Fis. 6 are erucal to > (ooo, aig racket, fore ee) Powder bed (sisted mei omogeey) Fig. 8. (3) Last decae of published papers on online sensing techiques and (6) extent of coverage of monitored signatures sing each specific sensing teenique. demonstrate extensive academic resonance, being the mest cited and presenting the highest H-index. Lower scientific relevance is associated ‘with multispectral approsches, while XR and sonic or ultrasonic (i, acoustic) sensing techniques gather certain interest bu are sil far rom visible-NIR and NIRAR ‘On the right side (Ng. 8b), a radar graph visually represents the ‘numberof covered signatures by each specific sensing technique, based ‘nthe information summarized in blo 5. Again, Visibe-NIR and NI IR sensors demonstrate extensive coverage, encompassing a wide range of signatures (ie, NIRAR for met pool and visibe-NIR for track and powder bed). Regarding ambient signatures, no scientific papers have ‘been retrieved addressing this aspect ‘rom a sensing perspective. Based on the above observations, itis evident that Visible-NIR is frequently employed in conjunction with NILIR systems (laser wave length and pool thermal signature), as discussed in Scion 5 (See Section (data variety), In the following Sections 4.21.1-4.2.1.5 we provide # detaled re. view of the retrieved papers for sensing techniques dived inthe ident fed clases and subasses (Vale 9). 42.1.1, XR, Ultrafast in-situ realtime XR imaging seems tobe the one ofthe most promising technique due to its relatively recent introduction and continuous grovth during the yeas. At the same time, compared to ‘other techniques suffer for less flexibility and more cosy and time expensive setup. As introduced by one of the first studies in [85], scientifically and technologically significant phenomens in metal L-PBF (eg melt pool dynamics, keyhole poresity, powder election and phase transformation) can be monitored via XR. On ths initial stage, al tional papers have been published aad an open architecture prototype of| a compact XR system was repored in [86]. step further in high-speed synchrotron XR imaging has been introduced by [62]. Similarly, eo the ‘previous papers they used a horizontal set-up (the laser had an angle of incidence of relative tothe sample surface normel) and 50 kHz frame tate. AS main result, the authors correlate the projective melt pool ge ‘ometties to laser absorption proposing the use ofa total backscattered light detection system for real time process control of Keyhole, melt pool's aspect ratio deviations, and instabilities [653). Furthermore, in [52] the authors propose an experiment on pulsed wave metal L-PBF to study the cavity and porosity patter formation (le, melt pool dy ‘namie), which isa wellknown issue in this operating mode (52]- Similar to [63] but more recent study has been reported by (87). By simulta ‘neous in-situ synchrotron XR (50 kHz frame rate), the authors directly ‘robe the interconnected fluid dynamics ofthe vapor jt formed by the laser and the depression it produces inthe melt pool demonstrating how "usstable keyhole is accompanied by a transition to chaotic flow in the plume [871 In the preset atile, the contribution on XR Computed ‘Tomography (XC) has not been included due tothe conventional off line (ie, post-process) inspection. This technique is only mentioned asa comparison (Le, ground truth) to online techniques, 42.12. Visble-NIR. Meanly dealing with visible or NIR spectra, digit Imaging is often associated with vision systems, which nowadays are ‘wemendously adopted both in the industrial and consumer industry using spatially resolved sensors such as Charge Coupled Devices (CCD) and Complementary Metal Oxide Semiconductor (CMOS) [38]. Due to the Mexdbilty, ow cost, and continuous improvement of visible/NIR maging techniques, they have been adopted for decades in laser-based sanufacturing and its level of adoption is at an early maturity stage. ‘Among the reviewed articles, in [59] a high-speed camera (20 kHz) was vurlized with telescopic lenses to monitor the interaction of laser pa rameters (Le, power and speed) and material signatures (.., melt pool ne spatter) In another paper, the authors presenta system based on & camera with 300 kHz fame rateas «compromise between the capability ‘ofeapturing the laser kinematies and the computational feasibility of in process image analysis (901. A similar approach was applied also in (91,82) on igh frame rate cameras (1-25 KH) some years ater. In the same year of [90], another step forward in the processing of image datasets generated from optical camera has been introduced by (92). 13 ‘their paper, the authors used a fixed (off-axis) field of view eamera "ranging in the 6-900 ki frame rate and propose a method to transform ‘the high-speed image daca such thatthe melt pools eppeer as viewed ‘through a comxially aligned optical setup. In 2019 a supervise learning technique (.e, Bayesian inference) was proposed asa solution for quas realtime (layer-wise) track (e3, pores/surface issues) control [94] Real time monitoring systems 10 enhance repeatability and quality control using collected videos to train CNNs with # semi-supervised F npieeat mode! are proposed by Yuan et al (136). These papers opened the era of ML adoption inthe optial imaging monitoring (and even forecasting) 35 demonstrated by the significant amount of optical sensing. and ML ‘contributions in the eceat years [95-105]. Papers specially focused ‘on spatter analysis were published in [79,106,107]. Other relevant po ‘pers, such as the one published in 2029 by (1051, report high-resolution init monitoring forthe identification of opical features correlated ‘withthe part density and mechanical properties. Similar papers based ‘on layer-ise optical control of defects on the layer surface were pro ‘posed in (54,109]. On the other han, in [110] the author proposed an ‘optical dimensional measuring method coding the surface profile layer ‘with structured light, which is recorded by a CCD camera, The point cloud coordinates output ere compared with the CAD model to detect, terrors in powder coating or layer thickness consolidation, Finally, we report another paper that proposes an online optical system based on industrial cameras to collect images and detect SLM powder bed spreading issues (eg., powder bed surface homogeneity) (971; for this specific application a low frame rate has been demonstrated tobe su ficient (0.023 kt). 42.13. NIRIR Heat transfer ia driving force of the L-PBF (111. The formation and dynamic behavior ofthe molten pool as well asthe liquid metal cooling and solidification have a direct impact on mrostrcture, residual stress, and deformation of components. Spatially integrated single channel sensors (e., photodiodes and pyrometers) have wide spread use in melt pool monitoring of advanced laser processing as well as cameras with special wavelength filters and frame rates [8]. A layer ‘wise approach based on relative surface temperature (i, thermal r ation) measurements was developed for subsurface defects identi cation (eg, lack of fusion) [112]. Another layer wise approach was Proposed by [113]. In this paper, the authors used a high-speed (100-250 H) thermal imaging system (Le, dual-wavelength imaging ‘pyrometer with temperature range 1500-2500 °C) to capture melt poo! temperature variations, Similarly, in [12] two-color pyrometry data sets were used to estimate instantaneous temperatures, melt pool of centations and aspect ratios [114]. Further application of pyrometers, adoption for melt pool temperature analysis can be retrieved. For ‘eample, [125] proposed melt pool temperature time series analysis via [ML for pores detection using a 100 k¥lz pyrometer as sensing source. ‘Additionally, temperature distribution in the sintering zane has been studied using a video camera, along with maximum surface temperature control inthe irradiation spot using a high-speed two-wavelengths py rometer {135 Other examples using IR camera and pyrometer applied for SLM process visualization and control can also be found varying the sampling rate ftom 2 (0 3 Kllz [15, In a more recent paper, another application of ML is proposed for IR data obtained from a thermal ‘camera (50 Hz) (116). Other applications on online monitoring of melt ‘pool geometry (Le, width/Shape) were carried out via TR camera with 58 He frame rate and a working range 600-3000 °C by (117 Similarly, 42 methodology for IR imaging feature extraction was proposed by [65,138]. In [18] the authors proposed an approach based on IR images Acquired with an offaxial IR camera (50 Ha frame rate) to detect possible deviations of plume from a stable behavior in AM zine powder processing (temperature range 100-500°C) [18]. Additional papers on processing thermal videos captured during the SLM of zine powder are reported by the same research group [69,119]. Furthermore, a Shor ‘Wave Infrared (SWIR) thermal camera (2.5 kz frame rate) measure ‘ments to compile voxel-based part representations and understand how the complesities in the thermal history affect part performance was proposed by 120]. The same author proposes an updated version ofthe Previous paper using SWIR thermal camera for porosity probebilty mapping [12]. Among the most recent papers, [122] detalls a design ‘and validation for an online insta monitoring system forthe detection ‘of some melt pool signatures (eg. temperature profile, temperature gradient, cooling rate) (127). Recently, another inprocess our of Manafcurg Frcs 108 (2023) 15-429 ‘thermography as an insitu monitoring tool wes presented by [123]. To conclude we also report a novel alteratve set-up for investigating the melt pool inthe cross section, based on high-speed (10 KH rame rate) thermograpbic camera mounted orthogonally to the seanned plane behind a glass [1241 42.14, Mulispecral. As opposed to the previously reviewed soch niques, multispectral emission spectrometry counts a few published papers. Despite Optical Emission Spectroscopy (OFS) has previousl ‘been implemented in laser welding processes for plume (eg, chemi species and temperature) and melt pool features (depth to width) as well as DED process monitoring, it seems that these approaches fr metal PBF are novel and at an early adoption stage [125,125]. In more de tals, in [125] the spectrometer i split into the Taser bear path of the SLM system o measure the visible light emitted from the melt pool size ane plume in-process at 14 Hz frame rte and correlated it with the melt ‘ool properties of samples. OES has been proved to be useful also for SSLM porosity monitoring. As described by another paper, optical em son signatures can be captured via in-situ multispectral photodetector sensor at 100 kHz frame rate [126]. Accordingly to the authors, the porosity-level within each layer ofa test part was quantified using XR ‘Computed Tomography (CT) ground truth to train @ Ml. mode! that take OFS data as input and prediet the percentage porosity-level in each layer. This approach is found to predict the porosity ona layer-by-layer basis with an accuracy of ~90 % in a computation time <0.5s [126] A ‘more recent paper stil based on ML prediction model proposes a dataset, obtained via an area-scan hyper fast (170 kz frame rate) byperspectral (Ge, from across the electromagnetic spectrum) camera a5 input to redier the surface roughness [27]. The hyperspectral images obtained from the process were labeled with the surface roughness as determined by a confocal microscope and used to tain a Convolutional Neural Network (CNN). Overall the results suggest that hyperspectal data does provide erucil information about the I-PBF process, and that hyper spectral imaging could potentially help establish on-line product qual feation [1271 42.15. Acoustic. Electromagnetic monitoring methods reviewed inthe ‘previous sections are common but are mainly limited to observing only thesurface ofthe AM build (except XR), Acouste sensing either inducing sonic or ultrasonic waves, s broadly adopted for nondestructive testing, ‘and monitoring internal features (e4 porosity) [128] due to the simplicity and low cost ofthe related sensors (he, microphones) [1 ‘Acoustic Bmision (AB) techniques are based on sonic or ultrasonic sensors, which receive waves tha relay back information about the in side and have been proposed for insitu process monitoring [135,137]. fist class of retrieved papers rely on Laser Ultrasonics (LU), which uses lasers to generate and detect ultrsonic waves and produce images known as B-seans and C-seans. For example, che authors of {125} pro ose an LU technique for porosity measurements. In their paper, the authors explore the current capability of the LU testing technique co detect subsurface defects in metal L-PBF in comparison to XR CT [125]. ‘Another paper based on LU testing in presented in (78, Here the authors ‘established a system to detect the surface defects of SLM samples that hhave a different surface roughness. The influences of the surface roughness on the LU signal-to-noise rato distribution and defect sizing accuracy were studied aswell [78]. Other papers focused on the ar-bors acoustic emission generated due to the SLM laser interaction to the ‘terial [129]. In their paper, the authors report the utilization of a ‘microphone mounted in the building chamber sampling acoustic data at 100 ki for porosity formation (Le, keyhole mode) monitoring. Sim larly [130] propose a method based onthe in-process acoustie signals ‘monitoring via 8 microphone. Finally, we recall the recent paper fom [151] in which an ultrasonic time of Might measurement monitoring technique s numerically and experimentally used to study the behavior of laser-induced melting pools including depth and width estimation as F npieeat our of Manafcurg Frcs 108 (2023) 15-429 ‘Table Pecinent retrieved papers concerning contr! strategy and thee features comparison, Ta) Hib anal bediedin ane ‘racial inlenernion snd eqns! of onal ymem and opel Ver iow Manuctrng Metology Tested NST GaneveCandocane oso (1) eon: Mel pool si/ ead car 40} Iligentcongesaon recture esd ‘Thee presertaten a cot ps (iow Grate, Powder ing, No Merce (01 Model Praise Conta OP) ‘Cempreon af cline PD ens in corto oftineM-hsnd prdisive No ik well as phase transition and other dynamics in real ime (231). To conclude, the most recent retrieved paper is about the application of transfer learning (Le, 2 paradigm where a model already trained on a similar tasks re-used with minimum training to accomplish anew task) 1 improve current ML generalization for online monitoring of different, ‘metal in SLM (152. In their paper, the authors propose the utilization fof AE spectrogram on a certain metal for capturing acoustic d: correlated to specific signatures (eg, pores, balling) and retain the already trained model to other metal via material specie spectrograms 032). 42.2. Online feedback strategies 4s highlighted by several authors, because of the LPBP process complexity (Le, over ity input parameters and as many observable signatures), actual feedback on AM machines is stl far from the i astral application [17,22]. To date, the L-PBF process is an open-loop ‘control system with several in-sit monitoring capabilites (Section 42.1). Unfortunately (and typically), a trial-and-error approach is adopted to set up experimental parameters, which are fixed throughout the entire build proces. These approaches ae insufficlent to ensure part ‘quality since parameters need to be dynamically adjusted in response to the underlying evolution of proces signatures [23,2], Despite the lack ‘of standard control protocols for L-PBF AM systems, in this section we report the restricted set of retrieved articles that propose an actual feedback control system and closed-loop strategy in detail ‘As reported in Section 2, most AM processes sar from a CAD fileand ‘the creation of digital position commands for galvo and power laser control asa series of x3-power entries stored into the G-Code. One the most used protocols for galvo control is the xy2-100 [159]. In his protocol, the digital position commands are packaged into 20-bit Packets and transmitted at 2 Miiz clock rate to the galvo digitl-to analog (D/A) receiver, converted to analog voltage to drive the galva ‘motor through a local Propontional-ntegral-Derivative (PID) loop. Similarly, the laser power on most commercial laser units can be controlled through an analog voltage input, where the digital power command can be transmitted and converted (D/A) as galvo control signal (159, According t the above, the authors of [129] propose = feedback control system for controlling the laser pover in an I-PBF process based on locally varying signatures (Le, relative proportion of solid or powder material near tothe melt pool) [138]. factor called the ‘Geometre Conductance Pactor (Gr) was ealeulated, and laser power was linesely scaled to the GCP theoughout the build modifying the power of the x--power array in ral ime for overhangs and edges. The algo rithm was implemented on the controller ofthe National Institute of Standards and Technology (NIST) open Additive Manufacturing ‘Metrology Testhed (AMMT) system, and the in-situ melt pool intensity ‘was measured via high-speed (10 KFiz rate frame) camera configured for co-axial melt pool monitoring (1391. As shown in Table © this contribution offers a practical and valuable application to the fed of contra "Another interesting application of feedback control can be found in the work of Adnan et. [10]. In thelr research, the authors proposed a eeack system based on a novel intelligent compensation architecture consisting of two systems (ie., System] and System2). Systemt is considered asthe primary control function and is designed for real ime control of the LPBF machine's discharging, coating, and polishing loops, lizing its fast and intuitive capabilities, On the other hand, ‘System? is developed as a secondary tuning function, employing a Melt ‘Pool Images (MPD-based approach with a CCD coavial camera and 2 hybrid model classifier based on ML tools, such as Convolutional Newral Networks (CNN) and Long Short-Term Memory (LSM). System2 aims to determine anc asses the appropriate thresholds and parameters of ‘System! based onthe ratios of different MP! classes in an offline manner (40) lastly, another paper on actual online control presented a theoretical approach and a valuable comparison between PID contol (i, the in dustry standard) and Model Predictive Coneol (MPC) in maintaining the required melt pool width and depth, controlling laser speed and power [041], One ofthe primary advantages of PID is its effectiveness when & ‘qualitative relationship between the desired output and control input i ‘known. After some paramcter tuning, control ear be conducted without the necessity of a physical model. For the I-PBF proces, if melt pool ‘width or depth is below the desired value increasing the laser power isa ‘well-established solution, Similarly, iflaser scanning speed is among the control inputs, reducing the scanning speed Is known to increase both ‘melt pool width and depth atthe same laser power level. Ina eases, PID controllers require diteet observation of controlled performance (0 derive an error tem for determining control ations. The capability for {nsitu monitoring, as reported in Section 4.2.1, indicates that melt pool features can be observed in realtime through various sensing ‘ech niques, with some constrains on the sampling rats inthe order of ki However, when multiple control inputs are used :o control the output response, here is «lack of mechanism to coordinate or optimize these inputs for achieving improved contral output. Furthermore, when can {rol objectives involve both observable and indirectly observable out pul, along with multiple control inputs, the effectiveness of the PID contol comes Into question [11]. For these reasons, the authors pro posed advanced MPC. While suck models are computationally expen sive, they are significantly more flexible in producing results under ferent setings. They can be conducted offine but are closely assoc ated with ML. Supervised Ml involves training a model on labeled datasets, where ‘the relationship between input data and output is known. The model learns from these labeled examples o predict outcomes for unseen data In the context of L-PBF, supervised learing algorithms can be employed tw classify defects, prediet temperature distributions, or estimate part ‘quality based on sensor date (Le, regression). On the other hand, un supervised ML involves taining a model on unlabeled datasess, where the structure and patterns in the data are not explcily known. By continuously monitoring the proces, such models ean detect abnormal conditions that might lead to defects or process failures, (24,203, 142,143]. Semi-supervised ML, as advanced technique, utilizes both labeled and unlabeled data, combining the benefits of both su persed and unsupervised approaches, Another advanced approach i= reinforcement learning, which incorporates an intermediary level of| Information into the algorithm, Ie is assumed that the taining data wil provide intermediate level of information about the true and false re sults, In this case, the system becomes “intelligent” fa trained env ronment based on rewards or punishments linked tothe accuracy ofthe results. Furthermore, hybridization represents the frontier of research F npieeat ‘here multiple models are adopted to optimize the overall system's performance, “leveraging” the stengths of diferent approaches to achieve enhanced results [24,104,141=143] ‘According to the above, MPC fest requires » physical model to cone the contol (eg, nite Element (FE) models that can pret the melt pool size (eg, length, width, and depth)) under different process parameter settings and assumed boundary conditions. Secondly, supervised ML algorithms are trained based on the generated datasets, from FE models. The third step is to further characterize the inherent ‘mode! bias under different process parameter settings, and to quantity ‘the uncontrollable process uncertainty (e.g, powder and laser property) Presented in the experiments. The final updated ML model can be ‘considered as a digital representation (or digital tn) of the physical process [11 'S. Results and discussion In Section 4, the four proposed drivers for an online contol of metal AM have been reviewed. We focused on the logical loop ofthe Deming ‘yele (aka PDCA) and considered the 95 papers resulting from a struc tured Iiterature review method proposed in Section 3 + Firstly, the online controllable input parameters (Le, 1st driver) ‘were reviewed in Seton 4.1.1 as the planning step (6, P) + Secondly the online observable ouput signatures (i.e, 2nd driver) ‘were treated in Section 4.1.2 as the result of the doing step (i.e, D) ‘+ Thirdly the online sensing techniques (Le, 3rd driver) were addressed in Section 4.2.1 as the monitoring or checking step (Le, C) + Finally, the few actual online feedback strategies (Le, th driver) retrieved were reported in Section 4.2.2 asthe eoreciveaetions oF act step (Le A). Online feedback strategies outlined in Section 4.2.2 se in-situ se sors identified Scio» 4.2.1) to collect data of online observable output signatures (defined in Seton 4.1.2) and adjust in real-time the related ‘online controllable input parameters (defined in Section 4.1.1). From the review of the previous sections, i is posible to derive a few final ‘considerations foreach driver and answer to the RO3: what are the Key {findings fom the lteranure? S.A. Input parameters/ouput signatures ‘The LPBF process has undergone extensive study and analysis, resulting in the establishment ofa universel hardware architecture (AM system or machine) and 2 specific physical process (selective fusion of ‘powder using a laser beam). Consequeatly, the input parameters and ‘output signatures ofthe process are well-defined and categorized Fig. 6 and Table 4 provide a summary ofthe first two drivers. However, its, ‘widely acknowledged in the literature that directly applying physics ‘heat transferor fuid dynamies equations to metal system is unfeasible due 2 various uncontrolled contingencies that impact each layer’s fabrication [14s]. For laser-based material processing, Dof. based ‘Response Surface Methodology (RSM) is commonly utilized for process development and determining processing parameter values that yield desired propecties to avoid anomalies [12]. While there as been sg nifleant research on experimentally and analytically understanding the correlations between input parameters and output signatures (eg, laser ppower/speed and melt pool signatures), the complexity ofthe overal ‘phenomena requires more advanced data driven approaches 5.2, Sensors Drawing on numerous papers providing an overview of sensing techniques. [3,21-25,83,84], this article presents a structured ura of Manafcarg Proce 108 (2023) 15-429 classification of sensors and their application in I-PBE, following the Innovative approach presented in Vig, 2 and detailed in Vig. 7. table 5 summarizes the corelation between sensors and process signatures, The reviewed sensors, extensively discussed in recent literature and under ‘going significant development in recent years, enable the monitoring of all significant (4, useful) proces output signatures. However, despite the progress in sensor development, implementing an ecient and structured sensor achitectre in AM industrial applications remains 2 challenging goal. While many papers focus on specific aspects of sensing, there is still lack of a general overview and a comprehensive ‘cost-benefit analysis. Based on the reviewed papers of Table 5, i i evidest that among various sensing echniques (Le, XR, visible-NIR, NIRIR, multispectral, acoustic), significant research attention Is directed towards visibie-NIR and NIRUR sensors, as depicted in Fig. 8a, This emphasis ean be attributed to their ability to offer cost effectiveness, simplified installation, and fast calibration processes, However, the primary reason lies in the extensive coverage of process signatures as depleted in ig. sb. 5.3. Control Despite the interesting papers that have been reviewed, the central message conveyed by this article isthe noticeable vod inthe literature concerning the practical application of closed online control loops in ‘etal AM, Out of the 217 reviewed papers, only one presents an actual implementation of online contol, and just two others offer compre Inensive theoretieal approaches that necessitate digital twin of the process (Vable 6}. This highlights the need for increased research attention in this crucial realm in dhe forthcoming yeas. As depleced in Pig. and serving as facilitator forthe concep of external offline loops, the closed-loop system not only provides real time control but also ‘generates invaluable data that can be stored for continuous process Improvement. Trough the analysis of historical sensor data and cor responding adjustments in process parameters, manufacturers can gain insights into proces trends and performance. This data-driven feedback loop enables refinements in ML. models and process strategies, leading to further optimization ofthe AM process overtime. 6. Outlook and open challenges ‘As discussed in the previous sections, researchers developed advanced sensing techniques for big data collection and maturity on sensors seems to be reached. On the other hang, Ml. is now the new ‘rontier of research for big data processing, which is sriecly related to sensors and control. The big data concept has never been so sulted a8 for thiscontextby defining the 3Vs (Velocity, Volume, Valet) as following lenges that address RQ4: what are the main future challenges? ‘Asamatterof fact, and as shoven in Section 4, the necessity of dealing with fast_growing (velocity), memoryintensive (volume), high dimensionality (variety) datasets is fostering the adoption of advanced 2M approaches to solve the problem of L-PBF data processing [15,14 Inthis context, data-driven techniques such as supervised, unsupervised ane reinforcement learning, are progressively assuming greater sgn cance in insita monitoring and process control tasks [18,112,143 6.1. Velocity chllenge Regardless ofthe type of sensor used, online control requires thatthe sensors have avery fast response time and a high degree of spatial res. ‘olution [21] Laser seanning speed in SLS/M are typically on the order of 100 to 1000 mms, while the laser focus area is on the onder of 10-100 sm, Any electromagnetic monitoring system must be equally capable of| reacting to these high scanning velocities and rapid melt pool dynamics, {in adétion to being able to resolve slight spatial variations (14. For F npieeat ‘example, phenomena such as spatter, plume formation, laser modula tion, and mel-poo! oscillations may require data acquisition rates ‘exceeding 10 ks. These constraints lead tothe utilization of relatively

    You might also like