딥러닝을 이용한 시계열데이터 군집화 (Clustering of Time Series Data using Deep Learning)

서지정보

· 발행기관 : 한국신뢰성학회
· 수록지 정보 : 신뢰성 응용연구 / 19권 / 2호 / 167 ~ 178페이지
· 저자명 : 윤동희, 김수민, 김도현

초록

Purpose: This paper presents the clustering results of time series multiple sensor data using deep neural networks based unsupervised learning algorithm without target variables.
Methods: Time series data collected from multiple sensors were clustered using two clustering algorithms based on deep learning: Deep Embedding Clustering (DEC) and Jointly Deep Embedding Clustering (JDEC). DEC and JDEC are designed based on the autoencoder and the convolutional neural network, which are representative neural network structures. They allow high- dimensional data to be represented by low-dimensional data and clustered based on their corresponding low-dimensional values.
Results: Two data sets, real time series data collected from manufacturing processes and simulated data, were used in the experiments. The simulated data’s performance was evaluated for accuracy, while the clustering performance of the real data was visually evaluated by mapping data and their clusters into a two-dimensional space. The experimental results show that the proposed methods were more accurate than K-means clustering.
Conclusion: Real time series data collected from manufacturing processes and simulated data were analyzed and meaningful clustering results were obtained. The proposed methods enabled the unsupervised learning of time series multiple sensor data without target variables by a deep neural network and showed good clustering performance.

영어초록

Purpose: This paper presents the clustering results of time series multiple sensor data using deep neural networks based unsupervised learning algorithm without target variables.
Methods: Time series data collected from multiple sensors were clustered using two clustering algorithms based on deep learning: Deep Embedding Clustering (DEC) and Jointly Deep Embedding Clustering (JDEC). DEC and JDEC are designed based on the autoencoder and the convolutional neural network, which are representative neural network structures. They allow high- dimensional data to be represented by low-dimensional data and clustered based on their corresponding low-dimensional values.
Results: Two data sets, real time series data collected from manufacturing processes and simulated data, were used in the experiments. The simulated data’s performance was evaluated for accuracy, while the clustering performance of the real data was visually evaluated by mapping data and their clusters into a two-dimensional space. The experimental results show that the proposed methods were more accurate than K-means clustering.
Conclusion: Real time series data collected from manufacturing processes and simulated data were analyzed and meaningful clustering results were obtained. The proposed methods enabled the unsupervised learning of time series multiple sensor data without target variables by a deep neural network and showed good clustering performance.

참고자료

· 없음