Chapter 8 Online melt pool monitoring using a deep transformer image processing solution
-
Javid Akhavan
Abstract
In the Industry 4.0 era, additive manufacturing (AM), particularly metal AM, has become a transformative technology due to its ability to efficiently produce complex geometries. Despite its potential, the industry still faces challenges in implementing robust real-time process monitoring algorithms. Recent advancements highlight that melt pool (MP) signatures during fabrication provide critical insights into process dynamics and quality. To capture this data, sensory systems such as high-speed camera-based vision modules are utilized for online monitoring. However, conventional analysis methods often struggle to process large volumes of data simultaneously. While traditional image processing (ImP) techniques offer a tunable approach, they present a trade-off between speed and accuracy, making them unsuitable for dynamic applications like MP monitoring. This work proposes the tunable deep transformer image processing (TDTIP) method, which integrates a hybrid convolutional auto-encoder (HCAE) architecture with a vision transformer (ViT) framework to meet the real-time demands of data-intensive monitoring. The model is trained in three phases: first, to replicate ImP algorithms with tunable features; second, to incorporate MP geometries; and third, to evaluate fabrication quality based on visual input and process parameters. TDTIP achieves over 94% estimation accuracy, with an R2 score exceeding 96% for the estimation and isolation of quality, geometry, and MP signatures. Additionally, the model processes data at an average rate of 500 frames per second (fps), a dramatic improvement over the time-consuming conventional methods. This significant reduction in processing time enables real-time vision-based monitoring for enhanced process and quality control.
Abstract
In the Industry 4.0 era, additive manufacturing (AM), particularly metal AM, has become a transformative technology due to its ability to efficiently produce complex geometries. Despite its potential, the industry still faces challenges in implementing robust real-time process monitoring algorithms. Recent advancements highlight that melt pool (MP) signatures during fabrication provide critical insights into process dynamics and quality. To capture this data, sensory systems such as high-speed camera-based vision modules are utilized for online monitoring. However, conventional analysis methods often struggle to process large volumes of data simultaneously. While traditional image processing (ImP) techniques offer a tunable approach, they present a trade-off between speed and accuracy, making them unsuitable for dynamic applications like MP monitoring. This work proposes the tunable deep transformer image processing (TDTIP) method, which integrates a hybrid convolutional auto-encoder (HCAE) architecture with a vision transformer (ViT) framework to meet the real-time demands of data-intensive monitoring. The model is trained in three phases: first, to replicate ImP algorithms with tunable features; second, to incorporate MP geometries; and third, to evaluate fabrication quality based on visual input and process parameters. TDTIP achieves over 94% estimation accuracy, with an R2 score exceeding 96% for the estimation and isolation of quality, geometry, and MP signatures. Additionally, the model processes data at an average rate of 500 frames per second (fps), a dramatic improvement over the time-consuming conventional methods. This significant reduction in processing time enables real-time vision-based monitoring for enhanced process and quality control.
Chapters in this book
- Frontmatter I
- Preface V
- Contents VII
-
Section: Image processing
- Chapter 1 Magnetic resonance image re-parameterization on real data 1
- Chapter 2 Denoising and gradient fusion for effective edge detection for noisy color images 17
- Chapter 3 Understanding driver attention to objects for ADASs: what do drivers see? 39
- Chapter 4 Image clustering enhanced with refined image classification 59
- Chapter 5 AI-powered framework for objective scoring of product design innovation 89
-
Section: Computer vision
- Chapter 6 Image inpainting using GAN transformerbased model 111
- Chapter 7 Enhanced image watermarking through cross-attention and noise-invariant domain learning 127
- Chapter 8 Online melt pool monitoring using a deep transformer image processing solution 153
- Chapter 9 Implementation of deep learning techniques on thermal image classification 173
- Chapter 10 Drishti: a generative AI-based application for gesture recognition and execution 203
-
Section: Pattern recognition
- Chapter 11 Exploring muzzle biometrics: a deep learning framework for noninvasive cattle recognition 239
- Chapter 12 Utilizing real-world data to develop a userindependent sensor-based human activity recognition system 253
- Index 273
Chapters in this book
- Frontmatter I
- Preface V
- Contents VII
-
Section: Image processing
- Chapter 1 Magnetic resonance image re-parameterization on real data 1
- Chapter 2 Denoising and gradient fusion for effective edge detection for noisy color images 17
- Chapter 3 Understanding driver attention to objects for ADASs: what do drivers see? 39
- Chapter 4 Image clustering enhanced with refined image classification 59
- Chapter 5 AI-powered framework for objective scoring of product design innovation 89
-
Section: Computer vision
- Chapter 6 Image inpainting using GAN transformerbased model 111
- Chapter 7 Enhanced image watermarking through cross-attention and noise-invariant domain learning 127
- Chapter 8 Online melt pool monitoring using a deep transformer image processing solution 153
- Chapter 9 Implementation of deep learning techniques on thermal image classification 173
- Chapter 10 Drishti: a generative AI-based application for gesture recognition and execution 203
-
Section: Pattern recognition
- Chapter 11 Exploring muzzle biometrics: a deep learning framework for noninvasive cattle recognition 239
- Chapter 12 Utilizing real-world data to develop a userindependent sensor-based human activity recognition system 253
- Index 273
