Multiscale Statistical Modeling Approach to Monitoring Mechanical Systems
Abstract
Signal processing for condition based maintenance and equipment monitoring has focused in recent years on non-stationary signal analysis using time-frequency representations of the signal. These representations are used to identify non-stationary events in the signal that indicate some change in the state of a structure or a machine. It is important to be able to reliably detect such changes in real time to do necessary preventive maintenance and also to minimize unnecessary maintenance. While transformations such as the Wigner-Ville, Gabor, and wavelet transforms are useful in highlighting time-frequency features of the signal, the application of such transforms to the monitoring problem requires additional for making decisions concerning the condition of the object being monitored. In particular, the interpretation of the transform coefficients in terms of physical events is essential to making such decisions. We develop a methodology for identifying the physical state of the object based on statistical models of the signals, which could comprise, for example, multiple outputs from devices such as accelerometers, strain sensors and acoustic emission sensors. Classification of machine states based on monitoring signals is performed by comparing likelihood scores for each machine state. We present examples of applying our system to various data, including damped sinusoids and noisy chirps, as a way of illustrating system performance for the case of transient monitoring signals. We compare our system to one which is trained using a DFT-based (non-time-frequency-based) representation (in particular, LPC coefficients) and show that our system exhibits both superior performance as well as greater robustness to noise in the signals.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1996
- Accession Number
- ADP010196
Entities
People
- Kenneth C. Chou