Rolling element bearings consist one of the most widely used industrial machine elements, being the interface between the stationary and the rotating part of the machine. The capability to detect fast and accurately the existence and severity of a fault in an installation during operation is crucial as an unexpected machine failure can lead to unacceptably long maintenance stops. In this paper, a novel online semi-supervised method is introduced for bearing fault detection and classification. A novel feature extraction method based on the "Mathematical Morphology" and the "Instantaneous Frequencies Estimation via Subspace Invariance Properties of (Complex Shifted Morlet) Wavelet Structures (IFESIS)" method is proposed. Initially the raw signal is transformed using a morphological operator by its interaction with a flat Structural Element (SE). The shape modified signal is further processed using IFESIS. The true (measured) bearing defect characteristic frequencies are detected by locating the highest Singular Values (SVs) of the corresponding Singular Value Decomposition (SVD) problem. The values of their harmonics are further calculated. Groups of characteristic harmonics are used by IFESIS in order to extract the nonzero SVs which consist the novel features and are further used as inputs to a Support Vector Data Description (SVDD) classifier focusing towards early abnormal change detection. Data captured during normal operation conditions are used for the training of the SVDD model fitting a tight hypersphere around them. The distance of the test data features from the center of the hypersphere is used as a health indicator. In case an abnormality is detected, a bearing fault classification is performed in a second step. The effectiveness of the proposed online method is validated using two bearing fault experimental cases from the NASA Prognostics Data Repository. The proposed method not only detects successfully anomalies at the early stage of failures bay additionally classifies the faults.
展开▼