Machine Vibration Monitoring for Diagnostics through Hypothesis Testing

Abstract

Nowadays, the subject of machine diagnostics is gathering growing interest in the research field as switching from a programmed to a preventive maintenance regime based on the real health conditions (i.e., condition-based maintenance) can lead to great advantages both in terms of safety and costs. Nondestructive tests monitoring the state of health are fundamental for this purpose. An effective form of condition monitoring is that based on vibration (vibration monitoring), which exploits inexpensive accelerometers to perform machine diagnostics. In this work, statistics and hypothesis testing will be used to build a solid foundation for damage detection by recognition of patterns in a multivariate dataset which collects simple time features extracted from accelerometric measurements. In this regard, data from high-speed aeronautical bearings were analyzed. These were acquired on a test rig built by the Dynamic and Identification Research Group (DIRG) of the Department of Mechanical and Aerospace Engineering at Politecnico di Torino. The proposed strategy was to reduce the multivariate dataset to a single index which the health conditions can be determined. This dimensionality reduction was initially performed using Principal Component Analysis, which proved to be a lossy compression. Improvement was obtained via Fisher's Linear Discriminant Analysis, which finds the direction with maximum distance between the damaged and healthy indices. This method is still ineffective in highlighting phenomena that develop in directions orthogonal to the discriminant. Finally, a lossless compression was achieved using the Mahalanobis distance-based Novelty Indices, which was also able to compensate for possible latent confounding factors. Further, considerations about the confidence, the sensitivity, the curse of dimensionality, and the minimum number of samples were also tackled for ensuring statistical significance. The results obtained here were very good not only in terms of reduced amounts of missed and false alarms, but also considering the speed of the algorithms, their simplicity, and the full independence from human interaction, which make them suitable for real time implementation and integration in condition-based maintenance (CBM) regimes.

Key-Phase-Free Blade Tip-Timing for Nonstationary Test Conditions: An Improved Algorithm for the Vibration Monitoring of a SAFRAN Turbomachine from the Surveillance 9 International Conference Contest

“…To conclude, the here proposed improved methodology fosters the on-line condition monitoring of turbomachines and could be used for diagnostic purposes if an anomaly detection (e.g., [24][25][26][27][28][29][30]) is conducted on the vibration signal of the different blades.…”

Performance of Envelope Demodulation for Bearing Damage Detection on CWRU Accelerometric Data: Kurtogram and Traditional Indicators vs. Targeted a Posteriori Band Indicators

“…Future improvements and applications may involve the definition of confidence intervals for the indicators [34], the application on different machineries and datasets such as in [35,36], the compensation of non-stationary rotational speeds [37][38][39], and the analysis of the effect of the sampling rate on the results.…”

Diagnosis of Faulty Wind Turbine Bearings Using Tower Vibration Measurements

“…To guarantee the reliability of the method, many measurement points are necessary [32,33]. The above features have therefore been extracted on independent (no overlap) chunks of the original signals: each acquisition is divided in 100 sub-parts which the five features are computed.…”

“…The computation of the Mahalanobis Novelty Indices trained on healthy, normal condition data was selected for its intrinsic ability of compensating for linear and quasilinear confounding influences (i.e., environmental or operational variability) and for its robustness to noise [32,38]. In fact, training the algorithm corresponds to fitting a statistical model to the reference data (i.e., healthy, normal condition data): if the training data set covers the whole range of normal variability, any outlier can be then attributed to a non-normal condition.…”