This project, jointly conducted by POSTECH, Seoul National University, KAIST, and KETI, aims to develop core technologies for Manufacturing Foundation Models (MFM) to advance AI-driven autonomous manufacturing. The research focuses on building a large-scale multimodal foundation model capable of learning from time-series, image, and text data collected from diverse manufacturing processes. The project includes designing multi-resolution encoders and decoders, developing cross-modal latent space fusion and instruction-tuning algorithms, and establishing an MLOps/AIOps pipeline for end-to-end lifecycle management of foundation models. The MFM will be validated through real-world demonstrations across automotive, shipbuilding, machinery, steel, petrochemical, and semiconductor industries, targeting tasks such as fault diagnosis, quality inspection, process optimization, and predictive maintenance. Ultimately, the project seeks to build a manufacturing-specialized AI ecosystem, enabling scalable, explainable, and interoperable AI solutions that strengthen Korea’s leadership in autonomous manufacturing innovation.

This project focuses on developing domain knowledge–embedded artificial intelligence for managing data drift in rotating machinery fault diagnosis. The research aims to address the limitations of conventional AI models that degrade under changing operating conditions by integrating physical knowledge–based drift detection, domain-informed AI algorithms, and adaptive retraining techniques into a unified framework. The study develops algorithms capable of detecting distributional shifts in real time, learning robust features from limited data using domain knowledge filters, and continuously updating models through domain adaptation strategies. Experiments are conducted using a pump testbed to simulate data drift scenarios and validate the proposed framework. The final goal is to enable real-time drift-aware fault diagnosis, reducing downtime and enhancing the reliability, safety, and efficiency of industrial rotating machinery systems.

This project, conducted in collaboration with the Korea Automotive Technology Institute, Autect Group, Sungwoo Motors, and MTR, focuses on the development and demonstration of purpose-built electric vehicles using a design platform. Within this scope, the research specifically targets the development of algorithms for road diagnostics and patrol purpose-built vehicles (PBVs). The project involves creating anomaly classification algorithms and road diagnostic algorithms. The project includes developing a deep learning algorithm for anomaly detection based on video data, which measures similarity to a mugshot to determine if a person is an anomaly. This involves training models with augmented mugshot images and enhancing the detection model through advanced object detection algorithms. Additionally, road surface diagnostics are addressed by developing edge processing techniques for identifying potholes in images, and combining multiple image data to assess road damage. The project also focuses on improving point cloud data processing for road severity assessment, including upscaling techniques for higher resolution and correcting data distortions due to increased driving speed.

This project involved developing a deep learning framework (DL framework) designed to automatically select the optimal model and parameters for building the best learning model using various industrial data. The framework employs partial domain adaptation and self-supervised learning as its models and includes an automated model design capability. These models were applied to and validated with real industrial motor data. 

The project focuses on research related to Prognostics and Health Management (PHM) for aircraft and helicopters. The primary objective of the study was to diagnose anomalies and predict engine degradation using Flight Data Recorder (FDR) data. To reach this objective, an innovative AI-driven diagnostic framework is introduced, which utilizes raw flight data converted into a correlation matrix. The detection of anomalies begins with an autoencoder model that is trained on normal data within the transformed correlation matrix. Following this, significant failure types are identified by cross-referencing them with historical data recorded by mechanics. This methodology is validated through the use of actual aircraft flight data.