01 Daim Ntawv Qhia
Additive Manufacturing (AM), raws li cov kev taw qhia tseem ceeb ntawm kev tsim cov thev naus laus zis siab heev, qhia tau hais tias muaj txiaj ntsig zoo hauv kev tsim kho cov khoom siv hlau thiab kev tsim cov qauv tsim. Txawm li cas los xij, thaum lub sij hawm hlau AM txheej txheem, kev sib cuam tshuam ntawm laser thiab cov khoom yooj yim ua rau muaj qhov tsis xws luag xws li tawg thiab porosity vim lub zog nqus tsis txaus, txwv nws qhov siab -precision industrial application. Laser absorptivity, raws li qhov tseem ceeb parameter txuas laser lub zog tawm tswv yim thiab cov khoom teb, yog qhov tseem ceeb heev rau kev kov yeej lub fwj dej ntawm qhov tseeb kom muaj nuj nqis thiab tiag - lub sijhawm kwv yees. Laser absorptivity ncaj qha txiav txim siab qhov kub thiab txias ntawm lub pas dej yaj; Kev nqus dej ntau dhau tuaj yeem ua rau tawg, thaum qis dhau yuav ua rau tsis muaj - ntawm - fusion defects. Txhawm rau hais txog qhov no, kev kawm sib sib zog nqus algorithms tuaj yeem ua tswvcuab, siv lawv cov duab kos tsis muaj zog thiab muaj peev xwm tshem tawm cov duab. Siv nyob rau hauv situ synchronized X-ray imaging ntawm keyhole collapse thwmsim (nrog rau qhov sib piv ntsuas ntsuas absorptivity) raws li cov ntaub ntawv tseem ceeb, haum convolutional neural networks (ResNet-50, ConvNeXt-T), semantic segmentation qauv muaj zog (UNet) cov qauv kev kawm muaj zog (UNet) nrog geometrical yam ntxwv ntawm keyhole vau (qhov tob, nam piv, thiab lwm yam) thiab absorptivity. Qhov no tuaj yeem tsim cov qauv kev kwv yees tseeb ntawm 'X-ray duab rau laser absorptivity' (ob qho tib si kawg-mus rau-kawg thiab modular txoj hauv kev), ua kom muaj tseeb-lub sij hawm kom muaj nuj nqis ntawm laser absorptivity thiab muab cov ntaub ntawv kev txhawb nqa rau kev tswj cov pas dej da dej dynamics thiab txo qhov tsis xws luag, yog li kev ua haujlwm siab ntawm AM.
02 Phau Ntawv Qhia Tag Nrho
Daim ntawv no tsim kev nqus thiab segmentation datasets siv cov ntaub ntawv tau los ntawm ib qho nyob rau hauv -situ synchronous high-speed X-ray imaging system ntawm 32-ID-B beamline ntawm Advanced Photon Source (APS) ntawm ANL, suav nrog cov ntaub ntawv porridge depression nrog txheej txheej, uas tsis muaj cov hmoov nplej, thiab cov txheej txheem uas tsis muaj cov hmoov av. siv raws li qhov kawg - mus rau - kawg thiab modular txoj kev. Qhov kawg-rau- txoj kev kawg siv ob lub convolutional neural networks, ResNet-50 thiab ConvNeXt-T, kom tau txais cov yam ntxwv cuam tshuam ncaj qha los ntawm ua ntej-processed X-ray dluab, outputting absorption rate los ntawm ib txheej X, tag nrho kev sib txuas ua ntej- tau kawm hauv ImageNet uas qhia qhov ua tau zoo tshaj plaws, ua tiav qhov kev xeem poob ntawm 2.35 ± 0.35 thiab qhov ua yuam kev qhov nruab nrab tsawg dua 3.3% ntawm cov hmoov- dawb Ti-6Al-4V kuaj. Txoj kev modular ua ntej rho tawm cov duab geometric ntawm cov vapor depression (xws li qhov tob, cheeb tsam, thiab nam piv) siv tus qauv UNet semantic segmentation, ces kwv yees tus nqi nqus siv cov qauv regression classical xws li Random Forest; UNet tau ua tiav qhov kev ntsuam xyuas siab tshaj plaws ntawm kev sib tshuam ntawm kev sib koom ua ke (mIoU) ntawm 93.5% hauv ntau-cov ntaub ntawv (xws li Ti64, SS316, IN718) segmentation cov dej num, thiab Random Forest qauv muaj qhov kev xeem poob ntawm 3.30 ± 0.02. Ntawm lawv, qhov kawg-rau-kawg txoj kev yog siab automated thiab ceev inference, haum rau industrial real-time xyuas, tab sis nrog tsis muaj zog txhais lus thiab loj twv ua yuam kev rau conduction qauv (me me vapor depressions); cov qauv modular muaj kev txhais lus muaj zog (kev ntsuas qhov tseem ceeb los ntawm SHAP qhov tseem ceeb, qhia meej txog qhov sib piv, qhov tob, thiab cheeb tsam raws li cov yam ntxwv tseem ceeb), tab sis cia siab rau qhov tseeb segmentation, nrog kev txwv tsis pub siv hauv cov hmoov av vim muaj teeb meem hauv kev txheeb xyuas thaj tsam kev nyuaj siab.
Daim duab 03 qhia cov duab tsom xam.
Daim duab 1 nthuav qhia qhov kev kwv yees tau tshwm sim ntawm laser nqus tsis muaj hmoov txheej. Subfigures a thiab b siv qhov kawg -rau- kawg ResNet-50 qauv, uas tuaj yeem taug qab cov kev hloov pauv hauv laser nqus tus nqi thaum lub sij hawm scanning thiab tiam sis nyob rau hauv qhov chaw nres tsheb laser tob qhov tseem ceeb theem, tab sis muaj qhov yuam kev loj nyob rau hauv thawj ob theem ntawm lub laser nyob ruaj ruaj. Subfigures c thiab d siv qhov kawg - mus rau - kawg ConvNeXt-T qauv, nrog scanning laser scenario yuam kev tsawg dua 3%, thiab nws kuj muaj peev xwm ua tau tseeb kwv yees qhov ntiav qhov tob theem ntawm lub laser nyob ruaj ruaj, nrog deviations tsuas yog nyob rau hauv tsis muaj- theem kev nyuaj siab. Subfigures e thiab f siv txoj hauv kev modular (UNet + random hav zoov), nrog rau kev ua haujlwm hauv scanning laser ze rau qhov kawg - mus rau - txoj kev kawg; Txawm li cas los xij, nyob rau theem tsis muaj kev nyuaj siab ntawm lub laser nyob ruaj ruaj, qhov kev kwv yees yog segmented li 0 (qhov sib txawv loj heev), thiab qhov tseeb yuav txhim kho tom qab cov ntaub ntawv ntiav ntiav.
Daim duab 2 piav qhia txog kev kawm ua tau zoo ntawm cov qauv sib txawv, qhov kawg -rau-kawg ResNet-50 qauv ua ntej- tau kawm (ImageNet hnyav) txo cov naj npawb ntawm kev sib koom ua ke los ntawm 19% piv rau qhov pib pib nrog me ntsis hauv kev poob, qhov kawg {-} ConvNeXt-T qauv ua ntej- kev cob qhia ua rau 69% txo qis hauv kev sib koom ua ke thiab kev poob qis (kuaj poob los ntawm 76%), thaum ua ntej- kev cob qhia UNet segmentation qauv tsuas yog txo cov kev sib koom ua ke nrog tsawg kawg ntawm 16%. Daim duab no qhia tau meej meej tias ua ntej- kev cob qhia hnyav hnyav txhim kho qhov kev ua kom zoo ntawm qhov kawg- mus rau- cov qauv kawg (tshwj xeeb yog ConvNeXt-T) tab sis muaj kev txwv tsis pub muaj cov qauv segmentation, muab cov lus qhia tseem ceeb rau cov qauv kev cob qhia zoo xaiv.
Figure 3 presents explanations and error analysis centered on the ConvNeXt-T model, comprising three subfigures: Subfigure a shows the attention distribution at different convolution stages through Grad-CAM heatmaps, illustrating the transition from dispersed attention in shallow layers to focused attention on the core region of the steam depression in deep layers, confirming the effectiveness of the end-to-end model in autonomously extracting key features; Subfigure b uses a 40% laser absorption rate as the threshold (distinguishing between conduction mode and keyhole mode) to analyse that samples with an absorption rate >40% (hom keyhole) muaj qhov kev twv ua yuam kev tsuas yog 2.54, qhov piv txwv tsawg dua lossis sib npaug li 40% (kev coj ua hom) muaj qhov yuam kev ntawm 12.6, qhia txog qhov yuam kev tseem ceeb ntawm cov qauv hauv kev coj ua hom; Subfigure c, los ntawm kev sim laser static ntawm 94W (tsawg zog, conduction hom) thiab 106W (siab dua lub zog, hom keyhole), ntxiv txheeb xyuas tias tus qauv qhov kev kwv yees zoo sib xws rau cov nqi tiag tiag hauv hom keyhole tab sis tsis tuaj yeem ntes cov kev hloov pauv hauv kev coj ua, corroborating qhov kev tshawb pom ntawm subfigure b.
04 Xaus
Txoj kev tshawb no tsom mus rau qhov kev kwv yees tam sim ntawm laser absorptivity hauv hlau additive manufacturing. Raws li synchrotron X-ray imaging thiab integrating sphere radiation measurements, datasets of Ti-6Al-4V absorptivity without and with powder, as well as multi-khoom keyhole segmentation datasets, were builded. Ob txoj kev kawm tob tau npaj tseg: kawg-rau- kawg (ResNet-50, ConvNeXt-T) thiab modular (UNet + random hav zoov), ob qho tib si ua tiav qhov kev kwv yees siab nrog MAE<3.3%, among which the pre-trained ConvNeXt-T end-to-end model performed best (test loss 2.35±0.35). ImageNet pre-trained weights significantly improved the convergence speed and accuracy of end-to-end models (ConvNeXt-T convergence rounds reduced by 69%, test loss reduced by 76%). Fine-tuning with a small amount of powder-containing data (5%) can effectively adapt to industrial scenarios. The end-to-end method is suitable for industrial real-time monitoring, while the modular method (explicitly considering aspect ratio, depth, and area as key features) is suitable for academic research and offline analysis.
