Vibration signatures contain information regarding the health status of the machine components. One approach to assess the health of the components is to search systematically for a list of specific failure patterns, based on the physical specifications of the known components (e.g. the physical specifications of the bearings, the gearwheels or the shafts). It is possible to do so, since the manifestation of the possible failures in the vibration signature is known a priory. The problem is that such a list is not comprehensive, and may not cover all possible failures. The manifestation of some failure modes in the vibration signature may be less investigated or even unknown. In addition, when more than one component is malfunctioning, unexpected patterns may be generated. Anomaly detection tackles the more general problem: How can one determine that the vibration signatures indicate abnormal functioning when the specifics of the abnormal functioning or its manifestation in the vibration signatures are not known a priori? In essence, anomaly detection completes the diagnostics of the predefined failure modes. In many complex machines (e.g. turbofan engines), the task of anomaly detection is further complicated by the fact that changes in operating conditions influence the vibration sources and change the frequency and amplitude characteristics of the signals, making them non-stationary. Because of that, joint time-frequency representations of the signals are desired. This is different from other vibration based diagnostic techniques, which are designated for stationary signals, and often focus on either the time domain or the frequency domain. For the purpose of this article, we will refer as TFR (time-frequency representation) to all 3D representations which employ on one axis either time, or cycles, or RPM, and on the other axis either frequency, or order. The proposed method suggests a solution for anomaly detection by analysis of various TFRs of the vibration signals (primarily the RPM-order domain). In the first stage, TFRs of healthy machines are used to create a baseline. The TFRs can be obtained using various methods (Wigner-Ville, wavelets, STFT, etc). In the next stage, the distance TFR between the inspected recording and the baseline is computed. In the third stage, the distance TFR is analyzed and the exceptional regions in the TFR are found and characterized. A basic classification of the anomaly type is suggested. The different stages of analysis: creating baselines, computing the distance TFR, identifying the exception regions, are illustrated with actual data.
CITATION STYLE
Klein, R. (2013). A method for anomaly detection for non-stationary vibration signatures. In PHM 2013 - Proceedings of the Annual Conference of the Prognostics and Health Management Society 2013 (pp. 230–236). Prognostics and Health Management Society. https://doi.org/10.36001/phmconf.2013.v5i1.2250
Mendeley helps you to discover research relevant for your work.