Unsupervised Feature Selection Using Recursive k-Means Silhouette Elimination (RkSE): A Two-Scenario Case Study for Fault Classification of High-Dimensional Sensor Data

Abstract

Feature selection is a crucial step to overcome the curse of dimensionality problem in data mining. This work proposes Recursive k-means Silhouette Elimination (RkSE) as a new unsupervised feature selection algorithm to reduce dimensionality in univariate and multivariate time-series datasets. Where k-means clustering is applied recursively to select the cluster representative features, following a unique application of silhouette measure for each cluster and a user-defined threshold as the feature selection or elimination criteria. The proposed method is evaluated on a hydraulic test rig, multi sensor readings in two different fashions: (1) Reduce the dimensionality in a multivariate classification problem using various classifiers of different functionalities. (2) Classification of univariate data in a sliding window scenario, where RkSE is used as a window compression method, to reduce the window dimensionality by selecting the best time points in a sliding window. Moreover, the results are validated using 10-fold cross validation technique. As well as, compared to the results when the classification is pulled directly with no feature selection applied. Additionally, a new taxonomy for k-means based feature selection methods is proposed. The experimental results and observations in the two comprehensive experiments demonstrated in this work reveal the capabilities and accuracy of the proposed method.

Sensor and Component Fault Detection and Diagnosis for Hydraulic Machinery Integrating LSTM Autoencoder Detector and Diagnostic Classifiers

“…The comparison included the application of the chosen ML and DL approaches using different features extracted or selected via numerous feature extraction and selection methods, such as manually extracting time domain features for each sliding window, such as the mean, variance, standard deviation, and signal to noise ratio. PCA was also applied directly using the raw multivariate time domain sensor data without dimensionality reduction, and by using the Recursive k -Means Silhouette Elimination (R k SE) feature selection and dimensionality reduction algorithm [ 36 ]. Finally, the trained models and saved thresholds from each experiment could be easily used to achieve run-time predictions of new samples at real-time.…”

Sensor and Component Fault Detection and Diagnosis for Hydraulic Machinery Integrating LSTM Autoencoder Detector and Diagnostic Classifiers

“…The comparison includes the application of the chosen ML and DL approaches using different features extracted or selected via numerous feature extraction and selection methods such as, manually extracting time domain features for each sliding window such as the mean, variance, standard deviation and signal to noise ratio. And applying PCA directly using the raw multivariate time domain sensor data without dimensionality reduction, as well as using Recursive k-Means Silhouette Elimination (RkSE) feature selection and dimensionality reduction algorithm [20] . Finally, the trained models and saved thresholds from each experiment can be easily used to achieve run-time predictions of new samples at real-time.…”