PHM system enhancement through noise reduction and feature normalization

The accuracy and precision of diagnostic features in a Prognostics and Health Management (PHM) system depends on the feature´s sensitivity to not only signal quality or signal-to-noise ratio (SNR), but also failure modes and operating conditions. The data acquired in real applications are not only a measure of the direct response of the system of interest, but also unwanted noises or abnormal signals. Before extracting diagnostic features, it is therefore critical to reduce the noises and eliminate the abnormal data (i.e., outliers) that can contaminate the data and result in a distortion or reduction in the amount of accurate information that can be obtained. In this paper, the authors will review existing outlier detection methods, such as Distance Based (DB), DBSCAN, and Minimum Covariance Determinant (MCD), and highlight their application to identify the data points that should be discarded. While these outlier detection methods will detect and help remove outliers, there is still a need to eliminate or reduce random noise or unwanted signals. Therefore, the authors have also applied a number of noise reduction techniques to the raw signal to increase the SNR. Finally, the on-board sensors mounted on a system operating in extreme environments can experience dramatic shifts in outputs over short durations or transient events. Furthermore, most existing diagnostic features are quite sensitive to the system´s operating conditions (e.g., speed, load, or torque) and therefore directly using the features for the system health monitoring can cause erroneous results or high false alarm rates. To mitigate the influence of transient data, the authors employ a clustering method for operational mode detection, as well as a novel feature normalization technique on the operating conditions. In this paper, the authors will compare and evaluate the contributions of the afore-mentioned algorithms and techniques to rotating machinery health monitoring.