Open AccessRemaining Useful Life Prediction Using Enhanced Convolutional Neural Network on Multivariate Time Series Sensor Data
Author(s) -
Manassakan Sanayha,
Peerapon VATEEKUL
Publication year - 2019
Publication title -
walailak journal of science and technology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.146
H-Index - 15
eISSN - 2228-835X
pISSN - 1686-3933
DOI - 10.48048/wjst.2019.4144
Subject(s) - univariate , computer science , mean squared error , time series , dropout (neural networks) , convolutional neural network , multivariate statistics , artificial intelligence , benchmark (surveying) , data mining , gradient descent , machine learning , artificial neural network , reliability (semiconductor) , overfitting , statistics , power (physics) , mathematics , physics , geodesy , quantum mechanics , geography
All machines in power plants need high reliability and to be continuous run at all times in the production process. The Remaining Useful Life (RUL) prediction of machines is an estimation for planning maintenance activities in advance to save the cost of corrective and preventive maintenance. Most existing models analyze sensor data separately. This univariate analysis never considers the relationship between sensors and time simultaneously. In this paper, we applied a Convolutional Neural Network (CNN), which considered both dimensions of and sensors; a multivariate time series analysis. Furthermore, we applied many techniques to enhance the framework of deep learning, including dropout, L2 Regularization, and the Adaptive Gradient Descent (AdaGrad). For the experiment, we conducted our method and showed the performance in term of Root Mean Square Error (RMSE) on a standard benchmark and for real-case datasets.