https://scholar.gist.ac.kr/handle/local/7916 2025-12-08T05:03:43Z https://scholar.gist.ac.kr/handle/local/31980 Title: 제조업 현장에서 딥러닝 프로젝트 실무 워크플로우 제안 Author(s): Taeyul Kim Abstract: 제조업에서 인공지능(AI) 기술 도입은 생산 효율 제고, 불량률 감소, 예방 정비 구현을 목표 로 빠르게 확산되고 있다. 그러나 제조 공정의 복잡성, 이해관계자의 다양성, 데이터 보안 제약 등으로 인해 딥러닝 프로젝트는 예산 초과, 일정 지연, 노사 갈등에 직면하기 쉽다. 본 연구는 이 러한 위험 요인을 최소화하기 위해 (1) 전략적 투자 기획, (2) 신속 PoC, (3) 데이터·모델 개발, (4) 배포·운영, (5) 사후 성과 분석의 다섯 단계로 구성된 실무 지향 워크플로우를 제안하고, 각 단계 에 RACI(Responsible, Accountable, Consulted, Informed) 매트릭스를 적용하여 권한과 책임을 명확히 규정하였다. 특히, 전략적 투자 기획 단계에서는 이해관계자의 요구사항과 예산 제한사항 을 반영하여 현실성 높은 프로젝트 목표를 설정하고, 신속 PoC 단계에서는 초기 타당성 검증을 통해 기술적 위험을 사전에 최소화하였다. 데이터·모델 개발 단계에서는 현장의 실제 데이터를 수 집하고 체계적인 품질관리를 수행하여 모델의 정확성과 신뢰성을 높이는 방법을 제시한다. 배포· 운영 단계에서는 현장 상황에 맞는 인프라를 선택하고, CI/CD 기반 자동화 시스템을 구축하여 안 정성과 효율성을 강화하였다. 사후 성과 분석 단계에서는 지속적인 모니터링과 자동 재학습 루프 를 통해 운영 중 발생 가능한 모델 성능 저하에 적극적으로 대응하였다. 워크플로우의 실효성은 금속 표면 결함 검출 사례(YOLOv8 기반)를 통해 검증하였다. 본 연구는 노사 협상, 인력 재배치, 산업 기밀 보호 등 현장 현실을 System 내에 내재화함으로써 기존 기술 중심 가이드라인의 한계 를 보완하였다. 워크플로우는 공정 복잡도와 기업 규모에 상관없이 적용 가능하며, 제조업의 AI 전환을 가속화하고 프로젝트 성공률을 제고할 수 있는 재현 가능한 청사진을 제공한다. 구체적인 워크플로우 내용은 3장에서 각 단계(3.1.1, 3.2.1 등)의 세부 내용을 통해 확인할 수 있다.|The adoption of artificial intelligence (AI) technology in the manufacturing industry is rapidly increasing to enhance production efficiency, reduce defect rates, and implement predictive maintenance. However, due to the complexity of manufacturing processes, diversity of stakeholders, and data security constraints, deep learning projects often face risks such as budget overruns, schedule delays, and labor-management conflicts. To minimize these risks, this study proposes a practitioner-oriented workflow composed of five phases: (1) Strategic Investment Planning, (2) Rapid Proof of Concept (PoC), (3) Data and Model Development, (4) Deployment and Operations, and (5) Post-Deployment Performance Analysis. Each phase clearly defines roles and responsibilities using the RACI (Responsible, Accountable, Consulted, Informed) matrix. Specifically, in the Strategic Investment Planning phase, realistic project objectives are set by considering stakeholder requirements and budget limitations. The Rapid PoC phase mitigates technical risks through early feasibility verification. The Data and Model Development phase outlines methods for collecting real-world data from the field and performing systematic quality control to enhance model accuracy and reliability. In the Deployment and Operations phase, infrastructure tailored to on-site conditions is selected, and stability and efficiency are improved by establishing a CI/CD-based automated system. The Post-Deployment Performance Analysis phase actively responds to potential model performance degradation through continuous monitoring and automated retraining loops. The practical effectiveness of the proposed workflow is validated through a case study involving the detection of metal surface defects using the YOLOv8 model. This study integrates workplace realities such as labor negotiations, workforce reallocation, and protection of industrial secrets into the workflow, addressing limitations of traditional technology-centered guidelines. The proposed workflow is applicable regardless of process complexity or company size, offering a reproducible blueprint that can accelerate AI transformation in manufacturing and improve project success rates. Detailed workflow descriptions for each phase (e.g., sections 3.1.1, 3.2.1, etc.) can be found in Chapter 3. 2024-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19895 Title: Weighted Loss Function Utilizing SNR Information for Monaural Phase-Aware Speech Enhancement Author(s): Jungwon Park Abstract: Speech enhancement is a task that suppresses background noise to improve speech quality and clarity for robust automatic speech recognition (ASR). In recent studies, deep learning-based approaches, which train the model from large amounts of data, have been extensively investigated. Recent studies have been proposed to estimate clean speech in complex and time domain to avoid the difficulty of accurately estimating the phase. In the case of phase-aware speech enhancement techniques in the complex domain, they can be divided into two types: mapping-based [5, 8] and masking-based methods. In our study, we use two convolutional recurrent network (CRN)-based [6-8] models of monaural speech enhancement and propose method to modify conventional loss function. In addition, we demonstrate that our proposed method performs better than the existing method by assigning different weights based on the signal-to-noise ratio (SNR) value. 2021-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19894 Title: Weather Aware Data Cleaning with Denoising AutoEncoder for Solar Power Generation Estimation Author(s): Junyoung Song Abstract: This paper proposes a data cleaning technique and a prediction method for diagnosing anomalous data of solar power generation facilities and predicting power generation. Accurate solar power generation forecasting plays an important role in optimizing the grid integration of renewable energy sources. One of the main challenges in solar power generation forecasting is the need to identify the presence of anomalous data, which can be considered as bad data affected by weather conditions. The proposed approach utilizes an AutoEncoder (AE), a type of unsupervised neural network, to perform anomaly identification while considering relevant weather conditions. AEs encode input data into a low-dimensional latent space to effectively filter out anomalous data that appears to be noise and capture underlying patterns that are affected by weather variables. For an effective data refinement process, we utilize a Denoising AutoEncoder (DAE) that performs well even with noisy data, which contributes to improving the quality of the data used for solar power generation forecasting. An experimental evaluation was conducted on a real-world PV power generation dataset to compare the forecasting model performance before and after applying DAE to weather input data. The results show that the DAE-based weather-aware data cleaning technique can mitigate the impact of uncertainty in prediction performance caused by noisy data, making the solar power generation prediction model reliable and universally applicable. 2022-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19892 Title: WatchCap: Improving Scanning Efficiency in People with LV through Compensatory Head Movement Stimulation Author(s): Jo, Taewoo Abstract: Individuals with Low Vision (LV) frequently face challenges in scanning performance, which in turn complicates daily activities requiring visual recognition. Although those with peripheral vision loss can theoretically compensate for these scanning deficiencies using active head movements, few practical applications have sought to capitalize on this potential, especially during visual recognition tasks. In this paper, we present WatchCap, a novel device leveraging the hanger reflex phenomenon to naturally elicit head movements through stimulation feedback. Our user studies, conducted with both sighted individuals in a simulated environment and people with glaucoma-related peripheral vision loss, demonstrate that WatchCap’s scanning-contingent stimulation enhances visual exploration. This improvement is evidenced by the fixation and saccade-related features and positive feedback from participants, without causing discomfort to the users. This study highlights the promise of facilitated head movements to aid those with LV in visual recognition tasks. Critically, since WatchCap functions independently of predefined or task-specific cues, it has a wide scope of applicability, even in ambient task situations. This independence positions WatchCap to complement existing tools aimed at detailed visual information acquisition, allowing integration with existing tools and facilitating a comprehensive approach to assist individuals with LV. 2023-12-31T15:00:00Z