• 지구물리와물리탐사
• /
• 제25권2호
• /
• pp.71-84
• /
• 2022

탄성파 자료 취득 시 신호와 함께 기록되는 다양한 형태의 잡음은 탄성파 자료의 정확한 해석을 방해하는 요인으로 작용한다. 따라서 탄성파 자료의 잡음 제거는 탄성파 자료 처리 과정 중 필수적인 절차이므로 기계 학습을 포함한 다양한 방식의 잡음 제거 연구가 수행되고 있다. 본 연구에서는 비지도 학습 기반의 탄성파 잡음 제거 모델을 이용하여 중합 전 탄성파 자료의 잡음 제거를 수행하고자 하였으며 총 세 가지의 비지도 학습 기반 기계 학습 모델을 비교하였다. 세 가지의 비지도 학습 모델은 N2NUNET, PATCHUNET, DDUL로 각각 서로 다른 신경망 구조를 통해 정답 자료 없이 탄성파 잡음을 제거한다. 세 가지 모델들을 인공 합성 및 현장 중합 전 탄성파 자료에 적용하여 잡음을 제거한 후 그 결과를 정성적·정량적으로 분석하였으며, 분석 결과 세 가지 비지도 학습 모델 모두 인공 합성 및 현장 자료의 탄성파 잡음을 적절히 제거하였음을 확인하였다. 그 중 N2NUNET 모델이 가장 낮은 잡음 제거 성능을 보여주었으며, PATCHUNET과 DDUL은 거의 유사한 결과를 도출하였지만, DDUL이 정량적으로 근소한 우위를 보였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권3호
• /
• pp.1722-1737
• /
• 2019

We developed an insider threat detection model to be used by organizations that repeat tasks at regular intervals. The model identifies the best combination of different feature selection algorithms, unsupervised learning algorithms, and standard scores. We derive a model specifically optimized for the organization by evaluating each combination in terms of accuracy, AUC (Area Under the Curve), and TPR (True Positive Rate). In order to validate this model, a four-year log was applied to the system handling sensitive information from public institutions. In the research target system, the user log was analyzed monthly based on the fact that the business process is processed at a cycle of one year, and the roles are determined for each person in charge. In order to classify the behavior of a user as abnormal, the standard scores of each organization were calculated and classified as abnormal when they exceeded certain thresholds. Using this method, we proposed an optimized model for the organization and verified it.

• 한국정보과학회논문지:소프트웨어및응용
• /
• 제34권2호
• /
• pp.112-122
• /
• 2007

말뭉치를 이용하여 통계적으로 의미역 결정(semantic role labeling)을 하기 위해서는, 의미역을 태깅하는 작업이 필수적이다. 그러나 한국어의 경우 의미역이 태깅된 대량의 말뭉치를 구하기 힘들며, 이를 직접 구축하기 위해서는 많은 시간과 노력이 필요한 문제점이 있다. 본 논문에서는 비지도 학습의 하나인 self-training 알고리즘을 적용하여, 의미역이 태깅되지 않은 말뭉치로부터 의미역을 결정하는 방법을 제안한다. 이를 위해, 세종 용언 전자사전의 격틀 정보를 이용하여 자동으로 학습 말뭉치를 구축하였으며, 확률 모델을 적용하여 점진적으로 학습하였다. 그 결과, 4개의 부사격 조사에 대해 평균적으로 83.00%의 정확률을 보였다.

• 한국군사과학기술학회지
• /
• 제19권3호
• /
• pp.346-355
• /
• 2016

In this paper, we propose a method to analyze the clusters of RF threats emitting electrical signals based on collected signal variables in integrated electronic warfare environments. We first analyze the signal variables collected by an electronic warfare receiver, and construct a model based on variables showing the properties of threats. To visualize the distribution of RF threats and reversely identify them, we use k-means clustering algorithm and self-organizing map (SOM) algorithm, which are belonging to unsupervised learning techniques. Through the resulting model compiled by k-means clustering and SOM algorithms, the RF threats can be classified into one of the distribution of RF threats. In an experiment, we measure the accuracy of classification results using the algorithms, and verify the resulting model that could be used to visually recognize the distribution of RF threats.

• 한국컴퓨터정보학회논문지
• /
• 제24권11호
• /
• pp.41-49
• /
• 2019

본 연구는 비지도 기계학습 기술과 코퍼스의 각 단어를 이용하여 한국어 단어를 형태소 분석하는 언어 모델을 구축하는데 목적을 둔다. 그리고 이 언어 모델의 단어 형태소 분석의 결과와 언어 심리 실험결과에서 얻은 한국어 언어사용자의 단어 이해/판단 시간이 상관관계을 갖는지를 규명하고자 한다. 논문에서는 한국어 세종코퍼스를 언어 모델로 학습하여 형태소 분리 규칙을 통해 한국어 단어를 자동 분리하는데 발생하는 단어 정보량(즉, surprisal(놀라움) 정도)을 측정하여 실제 단어를 읽는데 걸리는 반응 시간과 상관이 있는지 분석하였다. 이를 위해 코퍼스에서 단어에 대한 형태 구조 정보를 파악하기 위해 Morfessor 알고리즘을 적용하여 단어의 하위 단위 분리와 관련한 문법/패턴을 추출하고 형태소를 분석하는 언어 모델이 예측하는 정보량과 반응 시간 사이의 상관관계를 알아보기 위하여 선형 혼합 회귀(linear mixed regression) 모형을 설계하였다. 제안된 비지도 기계학습의 언어 모델은 파생단어를 d-형태소로 분석해서 파생단어의 음절의 형태로 처리를 하였다. 파생단어를 처리하는 데 필요한 사람의 인지 노력의 양 즉, 판독 시간 효과가 실제로 형태소 분류하는 기계학습 모델에 의한 단어 처리/이해로부터 초래될 수 있는 놀라움과 상관함을 보여 주었다. 본 연구는 놀라움의 가설 즉, 놀라움 효과는 단어 읽기 또는 처리 인지 노력과 관련이 있다는 가설을 뒷받침함을 확인하였다.

• 스마트미디어저널
• /
• 제11권4호
• /
• pp.30-37
• /
• 2022

Disease threatens plant growth and recognizing the type of disease is essential to making a remedy. In recent years, deep learning has witnessed a significant improvement for this task, however, a large volume of labeled images is one of the requirements to get decent performance. But annotated images are difficult and expensive to obtain in the agricultural field. Therefore, designing an efficient and effective strategy is one of the challenges in this area with few labeled data. Transfer learning, assuming taking knowledge from a source domain to a target domain, is borrowed to address this issue and observed comparable results. However, current transfer learning strategies can be regarded as a supervised method as it hypothesizes that there are many labeled images in a source domain. In contrast, unsupervised transfer learning, using only images in a source domain, gives more convenience as collecting images is much easier than annotating. In this paper, we leverage unsupervised transfer learning to perform plant disease recognition, by which we achieve a better performance than supervised transfer learning in many cases. Besides, a vision transformer with a bigger model capacity than convolution is utilized to have a better-pretrained feature space. With the vision transformer-based unsupervised transfer learning, we achieve better results than current works in two datasets. Especially, we obtain 97.3% accuracy with only 30 training images for each class in the Plant Village dataset. We hope that our work can encourage the community to pay attention to vision transformer-based unsupervised transfer learning in the agricultural field when with few labeled images.

• Geomechanics and Engineering
• /
• 제30권2호
• /
• pp.107-121
• /
• 2022

Coal pillar assessment is of broad importance to underground engineering structure, as the pillar failure can lead to enormous disasters. Because of the highly non-linear correlation between the pillar failure and its influential attributes, conventional forecasting techniques cannot generate accurate outcomes. To approximate the complex behavior of coal pillar, this paper elucidates a new idea to forecast the underground coal pillar stability using combined unsupervised-supervised learning. In order to build a database of the study, a total of 90 patterns of pillar cases were collected from authentic engineering structures. A state-of-the art feature depletion method, t-distribution symmetric neighbor embedding (t-SNE) has been employed to reduce significance of actual data features. Consequently, an unsupervised machine learning technique K-mean clustering was followed to reassign the t-SNE dimensionality reduced data in order to compute the relative class of coal pillar cases. Following that, the reassign dataset was divided into two parts: 70 percent for training dataset and 30 percent for testing dataset, respectively. The accuracy of the predicted data was then examined using support vector classifier (SVC) model performance measures such as precision, recall, and f₁-score. As a result, the proposed model can be employed for properly predicting the pillar failure class in a variety of underground rock engineering projects.

• Smart Structures and Systems
• /
• 제30권6호
• /
• pp.557-569
• /
• 2022

Vibration-based structural health monitoring (SHM) is crucial for the dynamic maintenance of civil building structures to protect property security and the lives of the public. Analyzing these vibrations with modern artificial intelligence and deep learning (DL) methods is a new trend. This paper proposed an unsupervised deep learning method based on a convolutional autoencoder (CAE), which can overcome the limitations of conventional supervised deep learning. With the convolutional core applied to the DL network, the method can extract features self-adaptively and efficiently. The effectiveness of the method in detecting damage is then tested using a benchmark model. Thereafter, this method is used to detect damage and instant disaster events in a rubber bearing-isolated gymnasium structure. The results indicate that the method enables the CAE network to learn the intact vibrations, so as to distinguish between different damage states of the benchmark model, and the outcome meets the high-dimensional data distribution characteristics visualized by the t-SNE method. Besides, the CAE-based network trained with daily vibrations of the isolating layer in the gymnasium can precisely recover newly collected vibration and detect the occurrence of the ground motion. The proposed method is effective at identifying nonlinear variations in the dynamic responses and has the potential to be used for structural condition assessment and safety warning.

• 한국멀티미디어학회논문지
• /
• 제23권8호
• /
• pp.1059-1065
• /
• 2020

In hospitals, nurses are subjectively determining the urine status to check the kidneys and circulatory system of patients whose statuses are related to patients with kidney disease, critically ill patients, and nursing homes before and after surgery. To improve this problem, this paper proposes a urine spectrum analysis system which clusters urine test results based on a hybrid machine learning model consists of unsupervised learning and supervised learning. The proposed system clusters the spectral data using unsupervised learning in the first part, and classifies them using supervised learning in the second part. The results of the proposed urine spectrum analysis system using a mixed model are evaluated with the results of pure supervised learning. This paper is expected to provide better services than existing medical services to patients by solving the shortage of nurses, shortening of examination time, and subjective evaluation in hospitals.

• 전기전자학회논문지
• /
• 제2권1호
• /
• pp.24-33
• /
• 1998

Independent Component analysis (ICA) is a new statistical method for extracting statistically independent components from their linear instantaneous mixtures which are generated by an unknown linear generative model. The recognition model is learned in unsupervised manner so that the recovered signals by the recognition model become the possibly scaled estimates of original source signals. This paper addresses the neural learning approach to ICA. As recognition models a linear feedforward network and a linear feedback network are considered. Associated learning algorithms for both networks are derived from maximum likelihood and information-theoretic approaches, using natural Riemannian gradient [1]. Theoretical results are confirmed by extensive computer simulations.