https://scholarworks.unist.ac.kr/handle/201301/149 2026-05-14T17:13:36Z 2026-05-14T17:13:36Z Yu, Hyeonseop https://scholarworks.unist.ac.kr/handle/201301/91526 2026-04-23T08:48:48Z 2026-01-31T15:00:00Z

Title: Design of a Fault Prediction Platform for Industrial Air Compressors Author(s): Yu, Hyeonseop Abstract: Industrial air compressors serve as critical equipment in manufacturing environments, supplying compressed air essential for production processes. Unexpected failures in these systems can lead to severe operational disruptions, including production downtime, delivery delays, and increased emergency maintenance costs. in particular ,Small and medium-sized enterprises (SMEs), often lack access to high-cost diagnostic instruments and advanced monitoring infrastructure, resulting in continued reliance on periodic inspections. Under such conditions, gradual degradation frequently remains undetected, leading to unexpected failures or excessive preventive maintenance. This study aims to address these challenges by designing and implementing a lightweight fault-prediction platform for industrial air compressors using a low-cost ESP32 microcontroller (MCU) equipped with current, vibration, temperature, and pressure sensors. A total of ten sensors were installed on the target compressor, collecting data at one-second intervals. The collected signals were processed through averaging, absolute-value conversion, and offset correction to generate stable feature values. Based on these processed features, fault-related patterns associated with motor load, bearing condition, filter clogging, and lubrication state were evaluated using simple rule-based logic, enabling basic fault-signature detection without complex formulas or high-performance computing resources. In addition, multichannel sensor data were used to evaluate the applicability of data-driven fault-prediction algorithms, Recurrent Neural Network–Long Short-Term Memory (RNN-LSTM), including Support Vector Machine (SVM), and Multivariate Vector Autoregression (VAR). Each sensor time series was normalized using mean and standard deviation to eliminate scale differences, and both normal and abnormal intervals were used to assess model feasibility. Visualization of pre- and post-normalization time-series data confirmed that the sensor signals were adequately transformed into a machine-learning-ready format. The research methodology consists of four main stages (1) reviewing compressor structure, fault characteristics, and condition-based maintenance (CBM) literature to determine monitoring targets. (2) designing the hardware architecture—including ESP32, sensor modules, power unit, and interface circuits—and implementing firmware for one-second data acquisition and preprocessing. (3) computing intuitive feature values such as current mean, vibration magnitude, and representative temperature and pressure values, and establishing a rule-based diagnostic logic using thresholds and reference levels. (4) deploying the platform on an actual industrial compressor, collecting operational data, validating platform behavior against maintenance logs, and applying machine-learning models to the same dataset to explore integration potential. Experimental results show that the ESP32-based platform reliably performed one-second data acquisition, preprocessing, and status evaluation in a factory environment. Variations in key sensor values—such as current, pressure, vibration, and temperature—enabled identification of major fault indicators, including increased motor load, suspected filter blockage, and potential bearing degradation. Data-driven analysis using normalized sensor data revealed changes in prediction errors and classification boundaries in intervals similar to those detected by the rule-based logic, demonstrating future applicability for platform enhancement. The main contribution of this study lies in presenting a practical, low-cost, and easily deployable fault-prediction platform capable of real-time monitoring and basic anomaly detection without expensive instrumentation or complex deep-learning models. Furthermore, by experimenting with SVM-, RNN-LSTM-, and VAR-based analyses on the same dataset, the study highlights the potential integration between lightweight embedded platforms and data-driven prognostic algorithms, suggesting future expansion toward TinyML and lightweight deep-learning-based predictive maintenance. Keywords: air compressor, predictive maintenance, condition-based maintenance, ESP32, sensor data, fault prediction platform Major: Master Degree in Information & Communication Technology (ICT) Convergence

2026-01-31T15:00:00Z Jeong, Inseok https://scholarworks.unist.ac.kr/handle/201301/91525 2026-04-23T08:48:48Z 2026-01-31T15:00:00Z

Title: A Study on Detection Model for Energy Poverty Households using Smart Meter Big Data Training Author(s): Jeong, Inseok Abstract: As the intensity of heatwaves and cold waves escalates due to the climate crisis, energy is being redefined as a necessity for survival rather than a mere commodity. However, the current income-based energy welfare system has limitations in detecting 'behavioral energy poverty' households that voluntarily forgo cooling and heating due to economic burdens. Therefore, this study proposes a 'Virtual Sensing' model capable of identifying potential poverty groups solely based on household electricity consumption patterns using established smart meter (AMI) big data, without requiring income information. This study conducted a 'Dual-Track Framework' combining micro-level modeling and macro-level verification. In Phase 1, an unsupervised clustering model based on temperature sensitivity (β) was constructed using high-resolution household data from Tokyo. In Phase 2, the social validity of the model was verified by fusing heterogeneous big data (electricity, weather, poverty) from Seoul. The major findings are as follows. First, the micro-analysis identified a 'Flat Pattern' cluster where electricity consumption hardly increased despite rising temperatures (β ≈ 9.03\%). This cluster (Cluster 2) remained at a minimal consumption level with an average electricity consumption of 70.3 kWh, and the average number of household members was 1.37, showing patterns similar to the characteristics of single-person vulnerable households. Second, the macro-verification revealed a distinct negative correlation (𝑅 ≈ −0.53) between summer electricity sensitivity and regional poverty rates, showing significant associations with housing benefit rates and the ratio of elderly living alone. In particular, this study elucidated 'Seasonal Asymmetry,' where summer electricity patterns (cooling as a luxury good) reveal consumption disparities based on income levels more clearly than winter patterns (heating as a necessity). This study empirically demonstrated the feasibility of detecting hidden poverty groups through a data- driven methodology despite privacy constraints. The proposed model is expected to contribute to the realization of 'Active Welfare' based on data, such as real-time monitoring of households in crisis by local governments and precision targeting of energy vouchers. Keywords: Behavioral Energy Poverty, Smart Meter (AMI), Virtual Sensing, Temperature Sensitivity, Unsupervised Clustering, Seasonal Asymmetry Major: Master Degree in Information & Communication Technology (ICT) Convergence

2026-01-31T15:00:00Z Kwon, Jeongbin https://scholarworks.unist.ac.kr/handle/201301/91523 2026-04-23T08:48:46Z 2026-01-31T15:00:00Z

Title: Design Methods and Characteristics of Active EMI Choke Using a Custumized IC for Real 11kW On-Board Charger Author(s): Kwon, Jeongbin Abstract: This thesis presents a design methodology for an active electromagnetic interference (EMI) choke aimed at enhancing common-mode (CM) impedance through complex impedance modeling and loop gain stability analysis. The feedback impedance is derived from the measured characteristics of a conventional CM choke and EMI filter, and is optimized to achieve both impedance boosting and system stability over the target frequency range. To implement this function in a compact form, a customized integrated circuit (IC) was developed and embedded within an 18 mm × 20 mm printed circuit board, enabling cost-effective mass production. The proposed design was experimentally validated using an 11kW on-board charger (OBC). Results showed that the active EMI choke increased CM impedance by more than five times around 300 kHz. Conducted emission (CE) measurements demonstrated that a two-stage EMI filter incorporating the active choke achieved comparable or superior suppression performance to a conventional three-stage passive filter while reducing overall weight by approximately 38%. Furthermore, tests with various magnetic core materials revealed that nanocrystalline cores provided the most effective broadband impedance enhancement. Major: Master Degree in Information & Communication Technology (ICT) Convergence

2026-01-31T15:00:00Z Park, SangHyun https://scholarworks.unist.ac.kr/handle/201301/91524 2026-04-23T08:48:47Z 2026-01-31T15:00:00Z

Title: An Integrated Multi-Task Learning Model for Predictive Maintenance (PdM) in Industrial Equipment Author(s): Park, SangHyun Abstract: Abstract Predictive maintenance (PdM) is increasingly essential in modern manufacturing environments, where unexpected equipment failures can cause production loss, quality deviation, schedule disruption, and elevated operational risk. While many PdM studies focus on a single objective such as Remaining Useful Life (RUL) estimation, practical maintenance decisions often require complementary outputs, including near-term risk indication and interpretable health-stage information. To address this gap, this thesis presents an integrated multi-task learning framework that jointly performs RUL prediction, imminent-failure detection, and health-stage classification from multivariate sensor time-series data. To ensure reproducibility, the study specifies an end-to-end pipeline including leakage-safe unit-level splitting, consistent preprocessing, fixed-length window construction, and unified label generation. Inputs are transformed into 30-cycle windows with a stride of 1 cycle, and labels are assigned to the most recent time point of each window for real-time inference alignment. RUL targets are capped at 125 cycles, imminent failure is defined as RUL ≤ 30, and health-stage classification uses three stages: normal (RUL > 80), degradation (31–80), and imminent (≤ 30). Experiments on the NASA C-MAPSS turbofan engine dataset (FD002 and FD003) show that the proposed shared-encoder CNN–LSTM model learns a degradation-relevant representation that supports multiple PdM objectives within a single network. Quantitative evaluation demonstrates comparable or improved task-level performance versus single-task baselines, and time-axis case analysis further indicates that operational usability is strengthened when outputs evolve coherently as failure approaches (declining RUL, increasing risk, and consistent stage transitions). The thesis concludes with limitations and deployment considerations, including domain-dependent variability, class imbalance in near-failure regions, and the need for calibration and governance of threshold-based actions, and suggests future work on consistency-aware objectives and deployment-oriented decision policies. Keywords: predictive maintenance, remaining useful life, imminent detection, health-stage classification, multi-task learning, time-series deep learning, CNN–LSTM, output coherence Major: Master Degree in Information & Communication Technology (ICT) Convergence

2026-01-31T15:00:00Z