Contents of Projects


Pilot Study: An AI-driven Approach to Developing Environmental Test Specifications for the Durability of Automobile Parts

[2024.08.26 ~, Hyundai Motor Company]

Objective

Prediction of the Durability Status of Vehicle Components Under Extreme Driving Conditions.

The Need for an AI-Driven Method to Predict the Durability Status of Vehicle Components

  • The traditional approach (physics-based models) for predicting the durability of vehicle components is vulnerable to changes in external environmental factors.
  • Furthermore, the physics models are often highly dependent on the domain knowledge (e.g. connection structures and design) of vehicle components, making it difficult to extend across different vehicle models.
  • Addressing these issues, data-analytic and deep-learning techniques are highlighted for their potential, which can automatically identify/select target-relevant variables during vehicle operation and determine their complex relationships.

Methodology

We compare several state-of-the-art approaches in predictive models to estimate the vehicle component temperature given a sequence of multi-variable inputs representing driving conditions.


Development of TOSEL problem generation platform by using generative AI

[2024.10.01 ~ 2025.01.31, Korea Research Foundation]

Objective

The goal is to develop an English problem generation platform that can handle text, image, and speech data using generative ai, and conduct research to advance problem quality.

  • (Text) English Vocabulary Question Generation Framework

Using the Llama 3.1-8B model and GPT-API, this framework creates high-quality TOSEL-aligned questions through optimized prompts and staged processes. Post-processing ensures consistency.

  • (Image) Personalized Image and Question Generation Framework

Combines fixed-frame images from AuraFlow-v0.3 with Llama 3.1-8B-generated questions. The LLaVA model integrates text and visuals for personalized, multimodal content.

  • (Speech) Automated Framework for English Listening Tests

Segments and logs voice data using deep learning, then automatically merges it to meet listening test standards.


Artificial intelligence(LLM)-based Question Generation Model Research

[2024.09.01 ~ 2025.08.31, EDU TOSEL]

Objective

The goal is to leverage LLM to generate different types of English problems and advance problem quality.

  • The Need for Auto-Generation of Multiple-Choice Questions (MCQs)

Generating multiple questions for exams by instructors or teachers consumes a significant amount of time. It is very challenging to generate questions that reflect individual differences in skill levels.

  • Existing Issues with Auto-Generation of Multiple-Choice Questions

When generating questions using language models, the types are fixed during training and generation, leading to a lack of flexibility in question formats. Questions are generated without reflecting the difficulty level based on the examinee’s proficiency.

  • Method

We study problem quality advancement without constraints from all methodologies, including Prompt Engineering, RAG, and Fine-Tuning.


Battlefield Situation and Data Simulation Technology

[2021-02-16 ~ 2022-11-30, National Defense Research Institute]

Objective

Develop a deep learning model that can generate unstructured simulation reports using structured simulation battlefield data.

  • Recently, many studies have been conducted to develop an intelligent battlefield recognition system in the military.
  • However, it is difficult to collect data for training.
  • In particular, there is a limitation in directly annotating reports generated through war games in the military. Therefore, it is intended to develop a deep learning model that generates a report from log data generated through war game simulation.

Data

Develop a deep learning model that can generate unstructured simulation reports using structured simulation battlefield data.

  • Recently, many studies have been conducted to develop an intelligent battlefield recognition system in the military.
https://univ20.com/11324

Related Work

1.Table-to-text Generation by Structure-aware Seq2seq Learning

  • In the encoding phase, they utilized the value-field as well as the postion, and proposed a modified encoder, the field-gate encoder.
  • In the decoding phase, Dual attention multiplied by word-level attention and field-level attention significantly improved performance.
Liu, T., Wang, K., Sha, L., Chang, B., & Sui, Z. (2018, April). Table-to-text generation by structure-aware seq2seq learning. In Thirty-Second AAAI Conference on Artificial Intelligence.

Proposed Method

Dual attention Seq2Seq with Copy mechanism

  • Previous studies have proposed replacing “unk” token with attention distribution as a solution to the OOV(out-of-vocabulary) problem.
  • However, this is a method that only applies if print out “unk” token.
  • Therefore, we propose a model applying Copy mechanism so that even words that are not in the vocabulary can be output.

Artificial Intelligence Research for Monomer Design

[2021-08-01 ~ 2023-07-31, Samyang]

Research Goal: to develop an artificial intelligence technology to design polymer with desired properties and create a structure-property database.

(1) Created a property database by collecting polymer data from public databases and literature.

  • Analyzed properties of each database and developed an auto-collecting tool for easy and fast collection. The auto-collecting tool can collect data that is updated later.

(2) Developed an artificial intelligence technology to analyze the relationship between molecule structures and properties.

  • Developed a new line notation for organic compounds and created a dataset that is required to train the artificial intelligence model.

(3) Explored the latent space to find a relationship between molecule structures and properties.

 

  • Exploring latent space that represents correlations within properties allows property analysis with low computational costs.
  • Attempted to apply VAE (Variational Auto Encoder)

(4) Generative Model-based Structure Prediction

  • Developed an artificial intelligence technology that predicts molecule structures with desired properties.

Development of AI-based Project Delay Risk Preview System

[2021-05-20 ~ 2021-10-31, Hyundai NGV, Hyundai Motor Company]

Objective

In the construction industry, the construction process is managed based on the completed construction in comparison to the planned construction. With the current construction management system, it is difficult to respond to construction delays in advance, because the construction process rate cannot be estimated quantitatively.

Data

Time Series Construction Data, Weather Forecast Data, Construction Budget Data, Extra-Budgetary Report Data, Subcontractor Data

Related Work

Similarity search and performance prediction of shield tunnels in operation through time series data mining (H.Zhu et al, Automation in Construction, 2020)

Proposed Method

In this paper, we developed an AI-based warning system that proactively predicts the delay in construction based on the completed construction of the past and additional data such as budget, subcontractor, and weather data.


Real-time Object Detection for Lithium Battery

[2021-06-28 ~ 2021-12-09, Korea Electronics Recycling Cooperative]

Objective

Lithium battery includes harmful metals (lead, mercury, etc.). Thus, collecting eco-friendly resources and managing hazardous materials are required during the discharge process. We propose a deep learning-based pipeline to select products containing the lithium battery among waste electrical and electronic products on a conveyor belt.

Data

Related Work

Related work: This study aims at the problem that mining conveyor belts are easily damaged under severe working conditions, based on the reclassification and definition of conveyor belt damage types. Conveyor belt damage is detected by the improved Yolov3 algorithm, which considers the impact of model scaling on the detection results [Zhang et al., 2021].

Zhang, Mengchao, et al. “Deep learning-based damage detection of mining conveyor belt.” Measurement 175 (2021): 109130.

Proposed Method

We trained Electronic and Electronical data based on YOLOv4 with CSPDarknet53 as the backbone and made the pipeline for detecting products containing the lithium battery on a conveyor belt. When waste electrical and electronic products move on a conveyor belt, a line scan camera detects the product, if the product is a lithium battery product. Then, the network signal the reject device to select the product.


Research on Artificial Intelligence Writing Technology Using Natural Language Processing

[2021-07-01 ~ 2021-12-31, Mania Mind (Research collaboration project) ]

Objective

Development of  Novel Writing Platform based on “user input” through learning of various novel genres.

Data

KoGPT2 fine tuning is performed using novel text data. In the case of Semantic Role Labeling, we use ETRI Semantic Role Labeling Corpus for training SRL model.

Related Work

KoGPT2 is a pretrained language model and optimized for sentence generation so that the next word in a given text can be well predicted. KoGPT2 is a transformer decoder language model that has been learned with more than 40GB of text to overcome insufficient Korean performance.

Proposed Method

It is a structure that combines Generate Layer for novel generation and SRL Layer for reflecting user input. When a sentence is entered, the following sentence is generated through KoGPT2 and the generated sentence is corrected through SRL layer.


AI consultation chatbot

[2021-07-01 ~ 2021-12-31, Korea University Anam Hospital (Research collaboration project)]

Objective

Development of a consultation AI chatbot for providing telemedicine counseling solutions

Data

EMR dataset was used for training the chatbot model.

Related Work

GPT-2 is a pretrained language model and optimized for sentence generation so that the next word in a given text can be well predicted. GPT-2 is a transformer decoder language model that has been learned with more than 40GB of text to overcome insufficient Korean performance.

Proposed Method

A Sequence consists of a list of questions and answers and a diagnostic name. It was used for fine-tuning GPT and predicting a diagnostic name. Therefore, it is possible to generate appropriate questions about the patient’s answers and finally predict the diagnostic name.