چكيده به لاتين
Recent studies have shown broad applications of traditional classification methods in dealing with industrial data, especially in fault detection and classification problems. These common methods mainly ignore time-series characteristics of data, which may lead to less accuracy and sensitivity in classification results. Using time-series classification methods is a new approach in dealing with fault detection and classification problems. In this thesis, we are using related machine learning algorithms for multivariate time-series classification of a specific dataset called Prognostics and Health Management, for the purpose of fault event detection and isolation in this industrial process. In the procedure, the dataset is preprocessed at first and all the related features, including novel statistical time-series related features, are extracted from the dataset. Various dimensionality reduction methods such as Principal Component Analysis and Random Forest are used to reduce the number of input variables. Before applying time-series classification algorithms, the prepared dataset is divided into training and testing sets in order to balance the classes in training dataset using resampling methods. Cross Validation method is also used for hyper-parameter tuning and finding the best algorithms for the dataset. Finally, time-series classification based machine learning methods such as K Nearest Neighbor, Linear Support Vector Machines, and Radial Basis Function Kernel Support Vector Machines are applied to the dataset for fault detection and classification.
