Deep Learning Approaches with Regularization, Generative Transformation, and Residual Learning in Biomedical and Material Sciences

Jaeyoon Kim 고려대학교 대학원 2025 국내박사

This thesis covers a variety of deep learning approaches with regularization, style-transfer generation, and residuals applied to biomedical and material sciences. The first study introduces a novel deep neural network architecture, deep learning-based cancer metastasis estimator (DeepCME), designed to predict breast cancer metastasis using gene expression. However, the problem of overfitting occurs frequently while training deep learning models using gene expression data because they contain a large number of genes, and the sample size is rather small. To address this problem, several regularization methods are implemented, such as L1 penalty, batch normalization, and dropout. DeepCME demonstrates superior performance, achieving the highest average area under curve scores compared with five baseline models. Additionally, significant genes associated with breast cancer metastasis are identified, highlighting the potential for cost-effective and time-efficient clinical applications. The second study introduces the combined transformation GAN (ctGAN), a style-transfer generative model designed to address existing challenges in the implementation of deep learning models for analyzing gene expression. Previous models have limited applicability for clinical purposes, but ctGAN enables the combined transformation of gene expression and survival data and improves survival analysis by augmenting data through style transformations. ctGAN demonstrates high plausibility in data generation based on distribution and clustering evaluations. The proposed method may enable predictions regarding the likelihood of a patient with breast cancer developing other types of cancer and responding differently to various treatment methods. The third study focuses on estimating Single-Parabolic Band (SPB) parameters in thermoelectric materials using residual-based ensemble model. Estimating SPB parameters is significant for understanding the band structure and predicting the theoretical performance of thermoelectric materials. However, the derivation of these band parameters involves complex calculations, including Fermi integrals, posing substantial challenges due to their computational complexity. To address this problem, this study proposes DeepSPB which is an ensemble model with deep neural network and machine learning model. Deep neural networks are trained to predict band parameters, including density of states effective mass, nondegenerate mobility, and reduced Fermi level. After that, an additional machine learning model leveraging residuals was implemented to adjust predictions by estimating error rates, effectively reducing the prediction error. The performance of the models was validated using pseudo data and experimental data for ZnSb decorated with Ag and Cu. The three methods described in this paper demonstrate considerable potential for application in various industries. The first two studies could improve metastatic cancer prediction and support cancer treatment and prevention in the medical field, benefiting public healthcare. The final study could facilitate material optimization in the semiconductor industry by analyzing band parameter variations with doping concentrations, reducing time and costs.

Machine learning을 이용한 시금치와 엇갈이배추 중 농약 잔류량 예측 모델 개발

김동주 충북대학교 일반대학원 2026 국내박사

This study was conducted to develop machine learning-based models for predicting time-dependent pesticide residue levels in spinach and Korean cabbage and to evaluate model accuracy by comparing predicted residues with actual residues through experimental validation. A total of 352 data points for spinach and 374 data points for Korean cabbage were collected for model construction. Using the spinach dataset, the Korean cabbage dataset, and an integrated dataset combining both crops, six algorithms, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting (GB), XGBoost (XG), and CatBoost (CB) were applied to develop predictive models. Among these models, RF, GB, XG, and CB demonstrated better predictive performance, as indicated by coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). For validation of model accuracy using actual field data, the test pesticides such as cyproconazole, dimpropyridaz, hexythiazox, and pyriproxyfen were tested. Field plots consisted of one and three-time application for the fungicide cyproconazole, and one and two-time application for the insecticides dimpropyridaz, hexythiazox, and pyriproxyfen, each arranged with three replications. Samples were collected at 0, 3, and 7 days after the final application. The pesticide residues in the samples were analyzed using LC-MS/MS following the quick, easy, cheap, effective, rugged, and safe (QuEChERS) extraction and clean-up. The limit of quantification was 0.01 mg/kg for all pesticides, calibration curves showed excellent linearity (R² = 0.9901– 0.9999), and recoveries ranged from 81.4–104.0% in spinach and 73.9– 103.9% in Korean cabbage. Mean error rates were calculated as the percent difference between actual and predicted residues at 0, 3, and 7 days after application. The error ranges for the spinach dataset and the integrated dataset were 19.9–44.2% and 24.3–49.2% for RF, 15.1–36.9% and 24.7– 50.5% for GB, 23.5–47.7% and 17.6–39.7% for XG, and 30.8–37.7% and 29.0–47.7% for CB, respectively, and those for the Korean cabbage dataset and the integrated dataset were 16.7–30.1% and 18.1–20.2% for RF, 25.6– 34.1% and 25.2–26.8% for GB, 15.4–17.0% and 16.1–18.3% for XG, and 50.5–88.2% and 36.9–54.6% for CB, respectively. Accordingly, the GB model from the spinach dataset showing 15.1% of mean error rate and the XG model from the Korean cabbage dataset exhibiting 15.4% of mean error rate were selected as the optimal models. Differences between experimentally derived half-lives and those calculated from the predicted residues were from 0.1 to 3.8 days for spinach and from 0.2 to 2.4 days for Korean cabbage. In contrast, the Dynamic Crop model, which is widely used internationally to predict pesticide residues in crops, showed substantially larger deviations for Korean cabbage, with errors ranging from 7.3 to 21.7 days. These findings demonstrate that machine learning models provide higher accuracy and reliability in predicting pesticide residue levels and half lives than the Dynamic Crop model and can serve as practical and efficient prediction tools for predicting pesticide residue levels in crops with limited residue data, such as minor crops. These results suggest that the machine learning-based model can be applied to assessing dissipation patterns, estimating half-lives, evaluating risks, and establishing maximum residue limits for pesticides.

Small object semantic segmentation using modified U-Net with morphological loss

심동우 연세대학교 대학원 2023 국내석사

Biometric Authentication with ECG and EEG based on Multi-scoring Residual Networks

김지훈 서울대학교 대학원 2020 국내박사

Biometric authentication based on electrocardiogram (ECG) and electroencephalogram (EEG) has witnessed remarkable progress in the past twenty years. Several anatomical and physiological patterns, such as fingerprints, faces, and irises, have been introduced and applied in some smart devices. However, those techniques are susceptible to forgery problems. Recently, a new generation of biometric authentication modalities, the biomedical signals that are typically used for clinical diagnostic purposes, has been suggested. Among the biomedical signals, ECG and EEG have demonstrated robust modality for biometric authentication. In this study, I developed ECG and EEG-based personal authentication algorithms using multi-scoring residual networks. First, I developed an ECG-based personal authentication algorithm. The ECG varies in waveform from person to person, and variations exist within individuals. In this study, I first constructed a database for the personal authentication test considering the characteristics of an ECG. Two versions of the ECG database have been built. The first database is a wired-ECG database. Three modified situations were measured with 105 healthy participants. The ECG was measured on two different days (time change), in two different postures (posture change), and before and after exercise (heart rate change owing to physical activity). The second database is a wrist-worn wireless ECG database. A dry electrode-based wearable and wireless ECG measuring device was developed and used to acquire signals from around 206 wrists; the ECG were measured at rest and before and after exercise. As a classifier for identifying individuals, ResNet, which is a deep learning neural network, is excellent for classifying images. One ResNet is inputted with one-dimensional ECG data, and the other ResNet utilized input images of the ECG signal through continuous wavelet transform. The fusion scores of the two parallel ResNets were scored using the Sugeno–Weber method, which achieved an equal error rate (EER) of less than 1% under a resting state and an EER of approximately 5% after exercise was performed. This was much better than those achieved in previous studies, especially after exercise, which were 2–20% lower than that achieved in this study. Second, I developed an EEG-based personal authentication algorithm. The EEG was analyzed based on the EEG of 100 test subjects registered in the Korea EEG Database. For the short EEG data of 10 s as one trial, 99 people were trained and tested for personal authentication. The residual network with the same structure used in the ECG authentication was used, and the measured time series EEG data were transformed into data in the frequency domain by Fourier transform and used as an input image. As a result, the accuracy of personal authentication improved to 95% when open/closed-eyes and to 99% when focusing on visual go/no-go tasks. The residual network-based EEG classifier presented in this study exhibited better results than other convolutional neural networks such as AlexNet and GoogLeNet. In addition, the accuracy was maximized through the score-fusion method, in parallel with the measured ECG. This is similar to an EER of 0.1–1%, which is now known as the fingerprint authentication accuracy. In this study, I proposed new authentication frameworks based on deep learning with highly accurate personal authentication performances and developed deep learning models to derive the optimal authentication performance. In the case of the ECG model, the accuracy of the change in the waveform was compared by using a wrist-wearable wearable device, and it provided robust and significantly enhanced results. In the case of the EEG model, the analysis accuracy could be useful when analyzing the authentication accuracy in the states of opening and closing the eyes and during visual tasks. Overall, this thesis attempts to provide the necessary deep learning-based algorithmic and practical framework for real-life deployment of the ECG and EEG signal in biometric recognition. 사물 인터넷의 발달과 핀테크 시장의 확장으로 기기가 개인을 인식하고 인증하는 기술에 대한 수요가 꾸준히 증가하고 있다. 기존의 사용자 인증 방식인 얼굴인식, 지문인식, 홍채인식 등과 같은 방법은 특징점이 외부에 노출 되어있어 위변조에 취약한 단점이 있다. 심전도, 뇌파와 같은 생물학적 지표인 생체 신호는 이에 비해 상대적으로 위조 및 변조에 강하여 차세대 인증 기술로 각광받고 있으나 개인 내의 변이성이 존재하여 정확도가 떨어진다는 단점이 있다. 한편, 최근 딥러닝 기술의 발전으로 패턴인식 분야에서 기존 분류 알고리즘보다 높은 성능의 알고리즘들이 각광을 받고 있다. 본 연구에서는 딥러닝 모델인 레스넷(ResNet)을 기반으로 멀티 스코어링 방식의 심전도 및 뇌파 기반 개인 인증 알고리즘을 개발하였다. 첫번째로 심전도 기반 개인 인증 알고리즘을 개발하였다. 심전도는 개인마다 파형이 조금씩 다르고 개인 내에서도 변이가 존재한다. 본 연구에서는 먼저 이러한 심전도의 특성을 고려하여 개인 인증 테스트를 위한 데이터베이스 구축을 하였다. 심전도 데이터베이스는 두 가지 버전으로 구축되었다. 첫번째로 유선 심전도 데이터베이스다. 105명의 건강한 참가자들로부터 세 가지 변형된 상황에 대해서 측정되었다. 즉 서로 다른 두 날 (시간 변화), 서로 다른 두 자세 (자세 변화), 그리고 운동 전후 (신체적 활동에 따른 심박 변화)의 상황에서 심전도가 측정되었다. 두번째 데이터베이스는 웨어러블 형태의 무선 심전도 데이터베이스다. 직접 개발한 손목착용형 건식 전극기반 무선 심전도 측정기기로 206명에 대하여 손목 부근에서 신호를 취득하였으며 휴식시, 운동 전후의 상황에서 심전도가 측정되었다. 개인을 식별하기 위한 분류기로는 이미지 분류에 뛰어난 딥러닝 신경망 중 하나인 레스넷을 이용하였고, 한 개의 레스넷에는 1차원 심전도 데이터를 입력으로 하고, 나머지 한 개의 레스넷에는 연속웨이블렛변환 기법을 거친 심전도 신호를 입력 이미지로 사용하였다. 두 개의 평행한 레스넷의 분류 점수를 Sugeno-weber 방법으로 융합한 결과, 휴식 상태에서 1% 미만의 동일 오류율(EER)을 보였고 운동 직후 심박이 증가한 상태에서는 약 5%의 EER을 보였다. 이는 선행 연구들에 비해 훨씬 더 좋은 결과였으며, 특히 운동 후의 결과는 그 전에 개발된 알고리즘에 의한 결과보다 EER이 2~20% 낮은 결과였다. 두번째로 뇌전도 기반 개인 인증 알고리즘을 개발하였다. 뇌파는 한국뇌파데이터베이스에 등록된 100명의 피시험자의 뇌파를 기반으로 분석을 진행하였으며 휴식상태의 눈을 뜨고 있을 때, 눈을 감고 있을 때, 그리고 시각 과업에 집중하고 있을 때의 세 가지 상태에 대해서 개인 인증을 각각 진행하였다. 비교적 짧은 10초의 뇌파 데이터를 하나의 인증시간으로 간주하여 신호취득이 불량한 1명을 제외한 99명에 대해 개인인증 테스트를 진행하였다. 심전도 인증에서 사용한 것과 같은 구조의 잔차 네트워크를 이용하였으며, 측정된 시계열 뇌파 데이터는 푸리에 변환으로 주파수 영역의 데이터로 변환하고 이를 이미지화하여 입력 이미지로 사용하였다. 그 결과 개인인증 정확도는 눈 개/폐시 95%, 시각 과업에 집중하고 있을 시는 99%까지 정확도가 향상되었다. 본 연구에서 제시한 잔차 네트워크를 이용한 뇌파 인증 분류기는 AlexNet이나 GoogLeNet과 같은 다른 컨볼루셔널 뉴럴 네트워크에 비해 좋은 결과를 보였다. 또한 같이 측정된 심전도와 병행하여 인증 점수 융합 (score-fusion) 방식을 통하여 정확도를 극대화할 수 있었다. 이는 현재 지문의 개인인증 정확도라고 알려진 0.1~1% 의 EER과 비슷한 성능이다. 본 연구에서는 위와 같이, 높은 정확도의 개인인증 성능을 갖는 딥러닝 기반의 새로운 인증 알고리즘을 제시하였고, 최적의 인증 성능을 도출할 수 있는 딥러닝 모델을 개발하였다. 심전도 모델의 경우 손목 착용형 웨어러블 기기를 이용하여 파형 변화시 인증 정확도를 비교하였고, 뇌전도 모델의 경우 눈 개폐시, 시각 과업시 인증 정확도를 분석하는 등 실제의 경우에 유용하게 활용될 수 있는 분석을 포함하였다는 관점에서 연구의 학술적 의의가 있다.

Hybrid Temperature-Emissivity Separation for LWIR Imaging via Numerical Estimation and Residual Learning

강민재 국립창원대학교 일반대학원 2026 국내석사

Long-Wave Infrared (LWIR) imaging is a critical sensing modality for industrial inspection, surveillance, and remote sensing due to its ability to operate independently of external illumination. However, accurate temperature retrieval from LWIR images is fundamentally challenged by the "thermal crossover" phenomenon, where the coupling between object temperature and surface emissivity creates spectral ambiguity. Conventional Temperature-Emissivity Separation (TES) methods typically rely on high-dimensional hyperspectral data, which incurs high computational costs and hardware complexity, or adopt constant emissivity assumptions that fail in heterogeneous material environments. To address these limitations, this paper proposes a novel hybrid TES framework designed for dual-band LWIR imaging systems. The proposed method synergizes physics-based numerical estimation with data-driven residual correction to achieve parameter retrieval robust against structural biases inherent in simplified radiative models with minimal spectral information. The framework consists of two sequential stages. First, an Adaptive Hybrid Numerical Solver computes an initial estimate of temperature and emissivity based on a simplified radiative transfer model. This solver integrates the Newton-Raphson method for rapid convergence with Brent's method for global stability, ensuring reliable initialization even under non-linear radiative conditions. Second, a Residual Compensation Network, designed based on the ConvNeXt architecture, corrects the structural errors introduced by the model simplifications. By learning the residual mapping between the initial numerical estimates and the ground truth, the network recovers physically accurate parameters without the need for complex multispectral measurements. Experimental results demonstrate that the proposed method significantly outperforms conventional single-band approaches and achieves accuracy comparable to multispectral techniques while maintaining high computational efficiency. This study presents a practical and effective solution for quantitative thermal analysis in dual-band LWIR sensing systems.

Deep Learning for Image Restoration and Robotic Vision

Du, Yixin ProQuest Dissertations & Theses West Virginia Univ 2019 해외박사(DDOD)

Traditional model-based approach requires the formulation of mathematical model, and the model often has limited performance. The quality of an image may degrade due to a variety of reasons: It could be the context of scene is affected by weather conditions such as haze, rain, and snow; It's also possible that there is some noise generated during image processing/transmission (e.g., artifacts generated during compression.). The goal of image restoration is to restore the image back to desirable quality both subjectively and objectively. Agricultural robotics is gaining interest these days since most agricultural works are lengthy and repetitive. Computer vision is crucial to robots especially the autonomous ones. However, it is challenging to have a precise mathematical model to describe the aforementioned problems. Compared with traditional approach, learning-based approach has an edge since it does not require any model to describe the problem. Moreover, learning-based approach now has the best-in-class performance on most of the vision problems such as image dehazing, super-resolution, and image recognition. In this dissertation, we address the problem of image restoration and robotic vision with deep learning. These two problems are highly related with each other from a unique network architecture perspective: It is essential to select appropriate networks when dealing with different problems. Specifically, we solve the problems of single image dehazing, High Efficiency Video Coding (HEVC) loop filtering and super-resolution, and computer vision for an autonomous robot. Our technical contributions are threefold: First, we propose to reformulate haze as a signal-dependent noise which allows us to uncover it by learning a structural residual. Based on our novel reformulation, we solve dehazing with recursive deep residual network and generative adversarial network which emphasizes on objective and perceptual quality, respectively. Second, we replace traditional filters in HEVC with a Convolutional Neural Network (CNN) filter. We show that our CNN filter could achieve 7% BD-rate saving when compared with traditional filters such as bilateral and deblocking filter. We also propose to incorporate a multi-scale CNN super-resolution module into HEVC. Such post-processing module could improve visual quality under extremely low bandwidth. Third, a transfer learning technique is implemented to support vision and autonomous decision making of a precision pollination robot. Good experimental results are reported with real-world data.

Track-Centric Iterative Learning for Global Trajectory Optimization in Autonomous Racing : x

강경태 울산과학기술원 Ulsan National Institute of Science and Te 2026 국내석사

Motion-Adaptive Few-Shot Fault Detection Method of Industrial Robot Gearboxes via Residual Convolutional Neural Network

오영탁 서울대학교 대학원 2020 국내석사

Nowadays, industrial robots are indispensable equipment for automated manufacturing processes because they can perform repetitive tasks with consistent precision and accuracy. However, when faults occur in the industrial robot, it can lead to the unexpected shutdown of the production line, which brings significant economic losses, so the fault detection is important. The gearbox, one of the main drivetrain components of an industrial robot, is often subjected to high torque loads, and faults occur frequently. When faults occur in the gearbox, the amplitude and frequency of the torque signal are modulated, which leads to changes in the characteristics of the torque signal. Although several previous studies have proposed fault detection methods for industrial robots using torque signals, it is still a challenge to extract fault-related features under various environmental and operating conditions and to detect faults in the complex motions used in industrial sites To overcome such difficulties, in this paper, we propose a novel motion-adaptive few-shot (MAFS) fault detection method of industrial robot gearboxes using torque ripples via a one-dimensional (1D) residual-convolutional neural network (Res-CNN) and binary-supervised domain adaptation (BSDA). The overall procedure of the proposed method is as follows. First, applying the moving average filtering to the torque signal to extract the data trend, and the torque ripples of the high-frequency band are obtained as a residual value between the original signal and the filtered signal. Second, classifying the state of pre-processed torque ripples under various operating and environmental conditions. It is shown that Res-CNN network 1) distinguishes small differences between normal and fault torque ripples effectively, and 2) focuses on important regions of the input data by the attention effect. Third, after constructing the Siamese network with a pre-trained network in the source domain, which consisted of simple motions, detecting the faults on the target domain, which consisted of complex motions through BSDA. As a result, 1) the similarities of the jointly shared physical mechanisms of torque ripples between simple and complex motions are learned, and 2) faults of the gearbox are adaptively detected while the industrial robot executes complex motions. The proposed method showed the most superior accuracy over other deep learning-based methods in few-shot conditions where only one cycle of each normal and fault data of complex motions is available. In addition, the transferable regions on the torque ripples after domain adaptation was highlighted using 1D guided grad-CAM. The effectiveness of the proposed method was validated with experimental data of multi-axial welding motions in constant and transient speed, which are commonly executed in real-industrial fields such as the automobile manufacturing line. Furthermore, it is expected that the proposed method is applicable to other types of motions, such as inspection, painting, assembly, and so on. The source code is available on my GitHub page of https://github.com/oyt9306/MAFS.

심층 학습 기반 리튬 이온 배터리 상태 진단 시스템

서동훈 충북대학교 2024 국내석사

LIB (Lithium Ion Battery) is an energy source characterized by high energy density and low self-discharge rate. It is used in various fields due to its high-output discharge and long usage cycle, and is especially actively applied as a power source for unmanned vehicles. As LIB continues to be used, a aging phenomenon occurs in which internal resistance increases. An aged LIBs have a reduced maximum charging capacity, which affects the maximum discharge current that can be output and the maximum operating time. Therefore, in order for a mobile vehicle to smoothly perform its mission, it is necessary to use a LIB with a guaranteed maximum charging capacity, and for this, a technology that can identify the maximum charging capacity of the LIB is required. Therefore, in this paper, we propose a deep learning-based LIB state diagnosis system to effectively identify the maximum charging capacity of the LIB. The proposed system is constructed using a deep neural network-based binary classification model, and extracts and synthesizes features from the input diagnostic data to output LIB status diagnosis results. Data for diagnosing the state is derived from time series data recorded during battery use, and vector-type diagnostic data using statistical variables was generated to effectively reflect the discharge and deterioration characteristics of the LIB. In addition, three states of LIB are defined considering the difficulty of the mission performed by the vehicles and the maximum charging capacity of LIB. To evaluate the performance of the proposed system, the model's diagnostic results were visualized using a confusion matrix, and the diagnostic performance for each state was analyzed in detail using recall and precision. Additionally, the feasibility of the proposed system was verified using random discharge experiment data that was not used for learning.

AI-based Heat Line Generation for Autonomous Manufacturing of Curved Plates

심경섭 국립창원대학교 일반대학원 2026 국내석사

The line heating process in the curved plate forming process has traditionally faced challenges in standardization and automation due to its complex thermo-mechanical nature and reliance on the tacit knowledge of skilled experts. To overcome these limitations, this paper proposes a deep learning-based heat line generation algorithm to realize the automation of the curved plate forming process. The overall concept is to learn the relationship between curved hull forming information and heat line patterns using an image-to-image translation-based deep learning technique. In a pre-processing step, to adopt an image-to-image translation techniques with fluent features for efficient training, point cloud data of the curved plates is converted into image-typed multi-channel data, and heat line position information is converted to the 2-channel-based image-typed data based on density and angle. The proposed network architecture integrates the encoder-decoder structure of U-Net with the Atrous Spatial Pyramid Pooling (ASPP) module of DeepLabv3+ to effectively extract geometric features across multiple scales. Finally, the heat lines are reconstructed by first estimating the center positions of each heat line from the predicted density map and then extracting the corresponding heat-line trajectories using the angle map. Quantitative and qualitative evaluation results demonstrate that the proposed model predicts a similar shape and value of density map in x and y domains compared to the ground truth. Implementation in a real curved hull forming process using hot working on actual steel plates with the heat lines generated by the proposed algorithm shows that the degree of completion is significantly improved toward the target curvature, with the completion rate increasing from 61.71% to 69.29%. These results confirm that the proposed deep learning model is applicable to real-world manufacturing environments and can contribute to enhancing the production efficiency of the curved hull forming process. 기존의 선상 가열 공정은 복잡한 열-기계적 특성과 작업자의 경험에 대한 높은 의존성으로 인해 공정의 표준화 및 자동화에 난항을 겪고 있다. 따라서, 본 연구는 이러한 한계를 극복하기 위해 곡가공 형성 과정의 자동화를 실현하기 위해 딥러닝 기반 가열선 생성 알고리즘을 제안한다. 전체적인 접근 방식은 곡면 형상 정보와 가열선의 패턴 간의 관계를 도출하기 위해 image-to-image 변환 기법을 사용하여 곡면의 형상 정보와 가열선의 패턴 사이 존재하는 관계를 학습하는 것이다. 전처리 단계에서 image-to-image 변환을 채택하고 풍부한 입출력 데이터를 정의하기 위해, 곡부재의 포인트 클라우드 데이터는 이미지 형상의 다중 채널 데이터로 변환되고, 가열선 정보는 밀도 및 각도 정보를 가지는 2채널 이미지 데이터로 정의된다. 제안된 네트워크는 3차원 형상 정보와 목표 형상 간의 기하학적 차이를 분석하기 위해 U-Net의 인코더-디코더 구조와 DeepLabv3+의 ASPP 모듈을 결합하여 설계된다. 마지막으로, 모델이 예측한 밀도 및 각도 정보로부터 각각의 가열선에 대해 예측한 중심점과 그에 대응하는 각도 정보를 통해 가열선이 재구성된다. 정량 및 정성 평가 결과는 제안된 모델이 x, y 도메인에서 정의된 밀도 정답 데이터와 유사한 분포를 예측함을 입증한다. 또한, 실제 강판을 이용한 열 가공 실험에서 제안된 알고리즘으로 생성된 가열선을 적용한 결과, 곡부재의 가공 완성도가 61.71%에서 69.29%로 유의미하게 향상됨을 보인다. 이러한 결과는 제안된 딥러닝 모델이 실제 제조 현장에 적용 가능하며, 곡가공 공정의 생산 효율성을 제고하는데 실질적으로 기여할 수 있음을 시사한다.