Our current research projects, outlined below, are dedicated to advancing metrology, sensor technology, and industrial AI techniques, with a primary focus on manufacturing and expanded applications in areas such as leak detection, fire monitoring, and bioengineering.
1. Real-time Defect Detection and Control of Laser Powder Bed Fusion using In-situ Sensors and Physics-Informed Machine Learning
One of the most promising additive manufacturing techniques is laser powder bed fusion (LPBF). An outstanding problem in LPBF is part quality issues caused by the complex and multi-dimensional relationship between its process parameters and part quality. Because of its layer-based fabrication nature, part quality is inconsistent even with the same material and process parameters, which requires discrete real-time process control with real-time metrology. This project addresses the knowledge gap of the real-time defect identification and the real-time process control to mitigate the defects. To bridge this gap, the project will address the following objectives: (1) Develop real-time measurement science for LPBF using multi-sensor, data fusion, and deep learning, (2) Develop a self-adaptive multi-dimensional correlation model between various defects and process parameters using deep learning, and (3) Develop a data-driven cyber manufacturing platform of LPBF for its real-time process control.
2. Image-Based Materials Characterization using Machine Learning (Sponsor: NSF)
A sufficient number of experimental datasets is crucial for training machine learning models. However, collecting large datasets in metals additive manufacturing is challenging due to experimental limitations. To address this, we developed a data augmentation technique using a denoising diffusion generative network. Our model successfully generates and interpolates cross-section images of metal parts from the experimental dataset. With the help of this data augmentation, we developed a machine learning model for characterizing pores and grain boundaries in cross-section images of metal parts. The model is generalizable and robust, capable of characterizing them in both 3D-printed and non-3D-printed materials. Validation results showed 90% accuracy in detecting pores and grains from cross-section images, even with unclear pores and grain boundaries due to non-optimized etching. The model also automatically extracted pore/grain morphology, including those areas and equivalent diameters. Additionally, we are developing a machine learning model to predict contact areas from cross-sectional images of metal parts. This model aims to characterize the evolution of the anisotropic distribution of contact areas.
3. Process Optimization of Sintering-Assisted Additive Manufacturing using Machine Learning (Sponsor: NSF)
Optimization of process parameters in sintering-assitive additive manufacturing is critical to achieve desired mechanical parameters of the final product. To achieve this complicated optimization, we are developing three machine learning models: a roller model, a binder jetting model, and a sintering model. These models will be integrated to optimize input parameters such as rolling speed and layer thickness for desired output conditions like densification. The database of actual and virtual sintering experiments assembled with the microstructure analysis data and the generative adversarial design model will be used to tackle the inverse design problem to optimize the process parameters of the sintering assistive additive manufacturing process. This multi-step machine learning approach can be applied to optimize a range of manufacturing processes.
4. Detection and Prediction of Water Pipe Leakage using Physics-Informed Machine Learning and Multi-Sensors (Sponsor: COMPA)
The limitations of current water pipe leak detection technologies are: (1) Lack of prediction technology: there is no technology available that can predict the location of leaks before they occur. (2) Difficulty in field application: existing leakage detection technologies using acoustic or vibration signals do not consider external noise, making it challenging to apply them in real-world conditions. (3) Lack of high-speed signal processing: there is a lack of automatic processing technology for high-speed sensor signals required for real-time leak monitoring. To address these challenges, the overarching goal of this project is to develop an Artificial Intelligence Sensor System (AISS) that utilizes AE/vibration sensors coupled with a physics-informed machine learning model. This system aims to detect and predict water pipe leak locations in real-time, offering a revolutionary approach to managing water infrastructure.
5. Dual Fringe Projection Moiré Profilometry (DFPMP) - New High-Resolution Full-Field 3D Measurement Technique
Emergence and advancement of high precision additive manufacturing (AM) techniques such as powder bed fusion require highly accurate full-field and real-time 3D measurements for monitoring process-induced defects and for ensuring the quality of printed products. However, most of the off-the-shelf 3D sensing techniques are not optimized for in-situ measurements for AM and the existing studies on topography measurement for AM failed to provide necessary resolution or field of view. To bridge this gap, the project will address the following objectives: (1) Develop a novel spatiotempotal FPP that applies simultaneous interferences in both spatial and temporal domains in FPP, (2) Apply multiplex color projection on the spatiotemporal FPP for real-time measurement, and (3) Study phase retrival from fringe patterns using deep learning.
6. Prompt Detection and Classification of Early-Stage Fires using Innovative and Cost-Effective Acoustic Field Sensors and Deep Learning Model.
The overarching goal of this project is to revolutionize fire detection by developing an innovative and cost-effective measurement technique to rapidly detect and classify fires at their inception, across any location in space, including hidden areas. This innovation seeks to address the limitations inherent in existing fire detection technologies, which often fail to identify and classify hazardous fires promptly, especially in their early stages or within concealed spaces. To achieve this goal, the project has three Objectives: (1) Develop pioneering sensor technology that leverages acoustic signals to detect the presence of fire within an area; (2) Develop a COMSOL simulation model for AF variation caused by fire; (3) Develop a deep learning model tailored for the accurate and rapid recognition and classification of early-stage hazardous fires through analysis of AF data.
7. Silent Speech Recognition using Wearable EMG/EEG Sensors and Machine Learning
Silent speech recognition (SSR) is a promising technology that transcends traditional speech interfaces by enabling communication without audible. This innovative approach leverages non-acoustic cues such as electroencephalogram (EEG) and electromyography (EMG) signals to convert into speech, providing a novel means of interaction for diverse applications in medical and military fields. However, current SSR studies have been limited to small vocabulary and specific applications, meaning that most studies are not transferrable and usable in other tasks. In addition, previous studies on SSR for sentences pose a strong restraint on data collection since manual pauses or signatures between words are required for segmenting words from continuous signals of sentences. To bridge this knowledge gap, we propose a moving-window based few-shot machine learning model that segments and classifies words from continuous EEG and EMG signals of sentences with small number of training data sets. To further increase the SSR accuracy and model robustness, we plan to integrate language models in the machine learning model.
8. Soft Tissue Characterization (STC) using Sub-microscale Tensile Testing
Characterizing the mechanical properties of biological tissues is a critical step in many medical and bioengineering fields such as regenerative medicine, robot-aided surgery, and mathematical modeling of biological systems. Although various tissue-characterization methods have been developed such as AFM, indentation, pipette aspiration, pressure myography, and elastography, there remains some limitations on accuracy and validity for measuring small and flexible soft tissues. To bridge this gap, the project will address the following objectives: (1) Develop a sub-microscale tensile testing technique to measure the mechanical propertoes of soft tissues and (2) Design a method for attaching the tissue samples to the clamps of the tensile tester and to avoid slipping of samples during testing.