Professor Byung-Dong Youn's Lab: System Integrity and Risk Management Lab

Head of the laboratory, Professor Byeong Dong Youn

Q1. What is the main research field of the laboratory, and what are some specific examples?
  Broadly speaking, there are three. The most active field is PHM (Prognostics & Health Management). Just as people maintain their health by responding to the symptoms of their bodies, this field measures the symptoms of industrial equipment in advance to prevent breakdowns. By sensing and analyzing various signals such as vibration, current, and temperature, you can obtain features, which are information related to the health of a facility. By observing features, you can gauge the health of your equipment and act. It is applied to automatic equipment robots, steam turbines, gas turbines, power transmission facilities, and can be applied to products such as automobiles and home appliances beyond industrial facilities.
Next is V&V (Statistical Model Verification & Validation), which models uncertainty in engineering product design based statistics and measures performance. For example, it is unclear what the distribution of load conditions in industrial facilities will actually be and what the physical properties of materials will be. These uncertain factors are reliably designed based statistics and  the performance is predicted.
The last is EH (Energy Harvesting). It is a technology that converts wasted vibration energy and thermal energy that exist in the natural world into electrical energy. Among them, the focus is piezoelectric EH that utilizes mechanical energy such as vibration/wave.

Q2. Do you have any special equipment or facilities in your laboratory? If not, is there any equipment that is often used outside?
Our laboratory is operating an industrial artificial intelligence manpower training industry group. We are conducting industrial site-tailored education for master's students from various laboratories within the Department of Mechanical Engineering, and are producing relevant workforce into the industry.
Another advantage is the large number of test beds. Failure data must be used to measure the health of industrial equipment. However, it is almost impossible to measure failure data of actual industrial equipment, so very good quality data can be obtained through scaled-down test beds. Given specific operating conditions, a test bed was created to acquire data when the equipment operates under those conditions. We have several test beds related to robots, motors, bearings, and turbines, and the permanent data obtained through them has created a better environment for research than other laboratories. In addition, there are many opportunities to deal with actual data as the research labs are in industry-academy cooperation with many companies.

Q3. Please tell us about examples of industrial application of the laboratory's technology.
We created a technology to prevent breakdowns by measuring the health of OHT (Overhead Hoist Transport) in Samsung Electronics' semiconductor factories. In a semiconductor factory, robots that move wafers frome process to another are transported from the ceiling. There are very complicated roads the ceiling and there are thousands of robots. In the complex transportation process, ife robot breaks down, all robots following it will stop. Downtime in a semiconductor factory causes huge losses, so we researched a technology to prevent it in advance. The health of the robot was continuously monitored, and if it was determined that there would be a problem, the robot was removed separately to avoid downtime. We measured the robot's current data to measure its soundness, and in this process, domain adaptation technology among deep learning techniques was used. This is a technique that adapts the network to the real world so that the network trained with training data works well with real data. This research is currently being used in Samsung Electronics' semiconductor plant.
Next, we are conducting a PHM research batteries for electric vehicles with Hyundai Motor Company. In order to improve the performance of electric vehicle batteries, research manufacturing by applying pressure to battery cells is in progress. The performance improvement of the battery was tested for each pressurization condition, and the distribution was statistically modeled, and the design was carried out based this. This study can be seen as a representative case study from the perspective of reliability-based design.
Lastly, we conducted research with KEPCO's IDPP (Intelligent Digital Power Plant). A leak is a phenomenon in which steam leaks due to a hole in the boiler pipe through which high-temperature, high-pressure water flows. This research is actually applied and used in several power plants in Korea.

Q4. I would appreciate it if you have something to say to undergraduate students who are thinking about graduate school or graduate students who are doing research.
Whether students fail or succeed, I would encourage them to try a lot and try this and that, regardless of the outcome. If you try a lot, there will be many failures. Don't be discouraged by the failure, learn what you did wrong in the failure and why you failed, and learn what to do so that you don't get the same result next time. If you repeat this process, there will be a moment when your growth rate is accelerated. What you try and fail when you're young may turn into nothing later. If you experience a lot when you are young, you will be able to see you who have grown up later.

---

Representative Researcher, Jinwook Lee

Q1. Please briefly introduce the selected representative research.
In the field of PHM, with the recent spread of IoT sensors and the development of deep learning technology, a lot of research deep learning-based fault diagnosis is being conducted. In order to learn a general supervised learning model it requires 1) data with a sufficient amount of labels and 2) training and test data that follow the same distribution, but it is difficult to obtain data for the conditions in the industrial field.
To solve the above problem, unsupervised domain adaptation techniques have been studied that transfer knowledge from the source domain with label information to perform tasks in the target domain without label information. However, the marginal distribution of each domain is well aligned, but the conditional distribution is not aligned. We created an algorithm that sorts conditional distributions by fusing maximum mean discordance and domain adversarial learning methods. In addition, the problems of instability in the learning process, slow convergence speed, and performance degradation in situations where the distance between domains is large in existing studies have been effectively improved. The high performance of the algorithm was confirmed through three different bearing data, and it seems to be able to contribute to solving the problem of lack of label data in the industrial field.
A semi-supervised domain adaptation technique that can utilize small label information even in the presence of some label information has also been studied in the laboratory. In the OHT failure diagnosis study conducted with Samsung Electronics, there was a problem that the data for each machine had a large deviation and the label data existed differently. Since the data for each unit is large, it is difficult to apply the data obtained bye type of robot to another type of robot even if it learns, and units with insufficient labels to proceed with unit-specific learning could not be learned. In order to solve this problem, we designed a domain adaptation algorithm using semi-supervised physically extracted features and corrected the severe data deviation for each unit to make an accurate diagnosis.

Q2. What was the most difficult part of the research and how did you overcome it?
When I first started my research, it was very difficult to implement the algorithm because I lacked a lot of knowledge about deep learning. We tried to implement the desired algorithm by using various deep learning books as well as the Internet. Also, there was a senior researcher who was conducting research a similar topic in the lab, so I was able to get a lot of help by asking questions. There were problems such as the algorithm implemented in this way not working as desired, but I carefully looked for problems and sought various solutions, and receiving continuous feedback from professors and members of the research team was very helpful.