Statistical Process Monitoring of Additive Firma convenzione Manufacturing via In-Situ Sensing Politecnico di Milano e Veneranda Fabbrica QPRC - June 13-15, 2017 del Duomo di Milano Marco Grasso and Bianca Maria Colosimo Aula Magna – Rettorato Department of Mechanical Engineering, Mercoledì 27 maggio 2015 marcoluigi.grasso@polimi.it, biancamaria.colosimo@polimi.it
Additive Manufacturing of metals Additive Manufacturing (AM) allows producing metal parts with innovative characteristics in terms of shape, surface properties, internal structure and overall value chain Topologically Functional surface patterns optimized shapes (e.g., healthcare) (e.g., aerospace) RUAG Lima Corporate GE Monolithic components (e.g., aerospace) Internal channels & conformal cooling Lightweight structures GE (e.g., medical & aerospace) 1 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Additive Manufacturing of metals How does it work? The laser powder bed fusion process Selective Laser Melting (SLM) A scanner displaces a laser beam along a predefined path to locally melt the metal powder, layer by layer Courtesy: Lawrence Livermore Laboratory • Laser beam diameter: 70 µm • Average powder particle diameter: 35 µm • Layer thickness: 50 µm 2 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Additive Manufacturing of metals The market situation and competitive scenario other • Leading sectors are aerospace and services healthcare • They are also the sectors with higher dental US Dollars TRL (e.g., GE aviation engine medical components, hip prosthesis, etc.) automotive aerospace 2023 2014 • Main AM system developers in EU • Merging & acquisitions involving big groups (e.g., Concept Laser & Arcam acquired by GE) • Many actors have impressive growth rates (e.g., 3D Systems: 52%; Arcam: 43%) • Technological competition mainly involves in-line monitoring and quality control 3 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Additive Manufacturing of metals The industrial barrier «The limited stability and repeatability of the process still represent a major barrier for the industrial breakthrough of metal AM systems» (Mani et al., 2015; Tapia and Elwany, 2014; Everton et al., 2016; Spears and Gold, 2016) AM process (SLM) High defective rates Residual Impurities & Geometrical (> 5 – 30%) Porosity stresses defects contaminations Current defective rates are not acceptable: Today, no commercial Expensive materials (e.g., titanium powders > 150 € /kg) system is able to Long processes (e.g., < 10 cm 3 /h) automatically detect Long/expensive trial-and-error inflates the time-to-market defects during the Stringent quality requirements (aerospace & healthcare) process 4 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Our background Statistical monitoring of product and process data Statistical monitoring of industrial processes for quick and reliable detection of out-of-control states and defects based on product and process data. Profile monitoring Surface monitoring Multi-fidelity data fusion product Controllable factors Colosimo et al. Colosimo et al. Colosimo et al. (2015), Input PROCESS (2008), JQT (2014), JQT Prec. Eng. Multi-sensor Profile monitoring Signal processing data fusion random, uncontrolled factors signals Grasso et al. (2016), Grasso et al. (2014), Grasso et al. (2016), JQT MSSP IJPR 5 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Our current research in metal Additive Manufacturing On-line monitoring via in-situ sensors Direct Energy Deposition ( DED ) Selective Laser Melting ( SLM ) Electron Beam powder & wire Melting ( EBM ) Prototype SLM Renishaw AM250 ARCAM A2 In-situ process monitoring and control (towards zero-defect metal AM) Image-based and multi-sensor statistical methods for on-line detection/localization of defects High-speed vision High-resolution low-speed vision Infrared vision Grasso et al. (2017), Grasso et al. (2016), Repossini et al. (2017) RCIM (under review) 6 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Case study Example of local over-heating in down-facing acute corners (AISI 316L steel) High-speed image acquisition (300 fps) Colosimo and Grasso (2017), Journal of Quality Technology (under review) Grasso et al. (2016), Journal of Manufacturing Science and Engineering 7 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach Corner A (no defect) Image stream Corner B (no defect) Y (N pixels) J frames Corner C (defect) X (M pixels) 350 frames of size 121 × 71 HOT-SPOT Intensity profiles over time (8bpp – scale: 0-255) 8 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach 𝓥 ∈ ℝ 𝐾×𝑁×𝑂 Image stream processing 𝓥 = {𝑽 1 , 𝑽 2 , … , 𝑽 𝐾 • Principal Component Analysis (PCA) applied to image data • No segmentation or edge detection operation needed Geospatial statistics & atmospheric science 9 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach Spatially weighted T-mode PCA (ST-PCA) Underlying idea : incorporating pixel spatial correlation into the projection entailed by the T- mode PCA to preserve the spatial depency and enhance the identification of local defects Weighted sample variance – covariance matrix: 1 𝐘 ∈ ℝ 𝑞×𝐾 is the data matrix (p=MxN pixels by J frames) 𝐲 𝑈 𝐗(𝐘 − 1 𝐓 = 𝑞−1 𝐘 − 1 𝐲 ) 𝐲 ∈ ℝ 1×𝐾 is the sample mean vector 𝟐 is a 𝑞 × 1 vector of ones 𝐗 ∈ ℝ 𝑞×𝑞 is the spatial weight matrix The (𝑙, ℎ) -th element of the matrix, 𝑥 𝑙,ℎ , quantifies the spatial dependency between the k-th and h-th pixels The matrix 𝐓 is a quadratic form whose decomposition into orthogonal components via eigenvector analysis has a closed analytical solution, being 𝐗 a symmetric weighting matrix 10 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach Spatially weighted T-mode PCA (ST-PCA) Use of Hotelling’s 𝑈 2 as a synthetic index to describe the information content along the most relevant components of the video image data within 𝐾 observed frames 𝑟 𝑨 𝑚,𝑗 2 where 𝜇 𝑘 is the l -th eigenvalue, (𝑛, 𝑜) are the pixel coordinates 𝑈 2 𝑛, 𝑜 = , 𝜇 𝑚 ( 𝑛 = 1, … , 𝑁, 𝑜 = 1, … , 𝑂) and 𝑟 is the number of retained PCs 𝑚=1 Hot-spot (corner) Normal 𝑈 2 (𝑛, 𝑜) Background melting Y (N pixels) J frames X (M pixels) 11 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach Spatially weighted T-mode PCA (ST-PCA) Two possible ways to iteratively update the ST-PCA as new frames become available Wold (1994), Gallagher et al. (1997), Li et al. Wang et al. (2005); De Ketelaere et al. (2015) (2000). 12 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Proposed approach Spatially weighted T-mode PCA (ST-PCA) Alarm rule based on k -means clustering of 𝑈 2 𝑛, 𝑜 • When process is IC : 𝑙 = 2 clusters are expected (background + normal melting) • When process is OOC : additional clusters correspond to defective areas (hot-spots) Automated selection of k based on sums of squared within-distances: k >2 ALARM (Zhao et al. 2009; Hastie et al. 2009) Cluster 3 (hot-spot) Cluster 2 No defect (IC) Hot-spot (OOC) (normal melting zone) Cluster 1 (background) 13 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Results Simulation analysis Simple T-mode PCA vs ST-PCA (Average Run Length – ARL) T-mode T-mode ST-PCA ST-PCA 14 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Hot-spot detection and localization in SLM Results Time of first Real case study Approach signal Example of T-mode PCA vs ST-PCA (frame index) OOC Scenario 1 Recursive No detection Average Simple T-mode PCA intensity Mov. window No detection 𝑘 = 201 T-mode Recursive 𝑘 = 198 PCA Mov. window 𝒌 = 𝟓𝟏 Recursive ST-PCA 𝒌 = 𝟓𝟏 Mov. window @ frame J=40 OOC Scenario 2 No detection Recursive 𝑘 = 144 Average intensity Mov. window No detection 𝑘 = 95 T-mode Recursive PCA Mov. window No detection 𝑘 = 94 Recursive ST-PCA 𝒌 = 𝟘𝟑 Mov. window ST-PCA OOC Scenario 3 Recursive No detection Average intensity 𝑘 = 173 Mov. window @ frame J=40 𝑘 = 169 T-mode Recursive Hot-spot detected PCA Mov. window 𝑘 = 168 𝑘 = 164 Recursive ST-PCA 𝒌 = 𝟐𝟔𝟒 Mov. window 15 marcoluigi.grasso@polimi.it; biancamaria.colosimo@polimi.it
Recommend
More recommend
