CAREER: Transforming Machine Learning Models Developed in Lab to Manufacturing Plants for In-Process Quality Prediction

  • Wang, Peng (PI)

Grants and Contracts Details


Overview: Over the past decade, the manufacturing community has been enthusiastically embracing Industry 4.0 techniques, such as Internet of Things (IoT), edge and cloud computing, and machine learning (ML), toward improved production visibility, planning, productivity, quality control, and safety. However, most ML solutions and tools are developed in lab environments and lack credibility for practical implementation in manufacturing plants. This discrepancy stems from the ML generalizability issue; ML models are sensitive to data, and there are major discrepancies between lab data and plant data, in terms of i) semantics or distribution, limited experimental tests conducted in labs vs. dynamic production conditions in plants; ii) amount and veracity, limited labeled lab data vs. big unlabeled pant data; and iii) modality, multi-modal advanced sensing data in labs vs. limited cost-effective sensing data in plants. Considering these discrepancies, making ML models developed in labs deployable to plants needs not a simple “transfer” but a substantial “transformation”, to achieve 1) good model generalizability, i.e., capable of both characterizing massive plant data from different conditions and fitting in different downstream application tasks; 2) enhanced understanding of process dynamics, i.e., being able to apply relationships between multi- modality sensing data learned from lab to generate virtual sensing data for improved process monitoring in plants; and 3) easy model adaptivity, easy model fine-tuning with limited labeled data to adapt to other application scenarios. Toward this transformation goal, this CAREER proposal aims at developing a novel transformer-based ML model architecture with configurable modules that allow self-supervised learning from massive unlabeled plant data, the interplay of task-agnostic data learning and task-specific model training, multi-modality data analysis and virtual data generation, and easy model adaptation. The proof- of-concept of the model will be deployed and evaluated in automotive robotic resistance spot welding (RSW) plants for process monitoring and quality prediction, through working with industrial collaborators. Intellectual Merit: To improve the credibility, applicability, and reliability of ML applications in plants, the ML models should meet the following expectations: 1) learning automatically how to extract features from massive unlabeled plant data toward characterizing the similarities or differences among data from different production conditions; 2) applying a generic data characterization structure to different downstream application tasks, considering the practical implementation constraints; 3) generating virtual sensing data for sensors physically unavailable in plants and leveraging data fusion for improved performance of quality prediction; 4) adapting easily models to future applications, with limited labeled data from experimental tests in lab needed for model tuning. To meet these expectations, the research will develop a transformer-based ML architecture that simultaneously realizes: i) task-agnostic self-supervised contrastive learning from massive plant data for robust data characterization and supervised model training to fit various downstream tasks; ii) normalizing flow for building one-to-one mapping between process sensing signals and artificial data generation; and iii) prompt ML model turning for transferring models between different application scenarios or plants. In addition, the proposed ML architecture is easily configurable and scalable for generic manufacturing applications to guide manufacturers in the U.S. in realizing their smart manufacturing plants. Broader Impacts: Manufacturing is projected to account for 27% of $14.4 trillion IoT market between 2013 and 2022. This CAREER project will boost the deployment of Industry 4.0 and ML techniques in manufacturing plants and accelerate the path to smart factories. Beyond the robotic welding plants to be covered in the project and the two major industrial collaborators, the developed generalizable ML architecture is easily expandable to a broad scope of manufacturing processes and will benefit the entire manufacturing sector. Especially, through working with the Kentucky Association of Manufacturers (KAM), the project outcomes will be disseminated to small-medium manufacturers in the Commonwealth-State of Kentucky, to help them realize digital transformation and grow their business competitiveness. Furthermore, novel theories developed in this work will impact the fundamental science community and have the potential to become widespread in the natural sciences, engineering, and healthcare applications, such as big data-supported health tracking systems. On the educational front, the research outcomes will be incorporated into curriculum development and high-school 360-hour research projects. Also, demo website and technical webinars will be held for manufacturers, to maximize the social awareness of the research outcomes.
Effective start/end date5/1/234/30/28


Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.