EXPERIMENTAL INVESTIGATION ON THE USE OF VIBRATION SIGNALS COMBINED WITH SUPERVISED CLASSIFICATION TO PREDICT RADIAL LOAD CONDITION IN ROLLER ELEMENT BEARINGS

In this work, vibration response of a rolling element bearing under the influence of static radial loading is investigated. Radial loading results in a periodically varying stiffness (or compliance) which causes a cyclic dynamic response of the bearing assembly even under perfect balancing and other operating load conditions. These loads cause high stresses to develop in bearing elements and may cause fatigue, cracks, and spalls that limits the life of these components. A special bearing test rig was designed and manufactured to apply varying levels of radial load and measure the vibration response of the loaded roller bearing. The test is focused on new bearings free from any faults or defects. The radial load is varied in steps and the vibration signal is collected and analyzed at each level for different rotor speeds. The spectral components are analyzed using Fast Fourier Transform (FFT) and time-frequency wavelet transform. Statistical techniques are applied to both the vibration signature obtained using a piezoelectric accelerometer sensor and the wavelet decomposed approximations and details of the original vibration signals. The statistical measures, wavelet approximation and details are first processed for feature set reduction since many of the features are highly correlated. This is done using three feature reduction and subset selection methods - ReliefF, Recursive Feature Extraction (RFE) and Multi-Cluster Feature Selection (MCFS). These features and the original extracted features are used as features to train two classifiers. The classification is used to estimate high and low thresholds for both radial load and running speed. The classifiers used are (1) radial-basis function support vector machine (RBF-SVM), and (2) k-nearest neighbor (kNN). Performance of machine learning algorithms depends on the training data and physical collected datasets are often limited to specific operating conditions, necessitating the use of training with many models using multi-fold cross-validated subsets. In this study we have used ten models using two-fold cross validation for training and validation. The classification results reported are average of these models. In limited experimentation, the RBF-SVM outperforms the kNN classifier and among the feature sets used, the ReliefF set seems marginally superior to the other sets. However, the accuracy, precision, and recall (combined as an F-score) of the original extracted feature set are better than the reduced feature sets; the downside being the relatively high run time in the training phase.