Fault detection in variable speed machinery: Statistical parameterization

Variable speed machinery presents a particular challenge to automated condition-monitoring systems; changes in speed have a strong relation to the vibration response collected by accelerometers—the effect of which may mask fault conditions in standard condition monitoring techniques. In order to account for the effects of this measurable variable, the vibration response will be segmented into speed bins with a small range of speed. The mean and covariance matrix for the feature vectors in each speed bin will be computed in order to derive a statistical novelty boundary for that bin. Each component of these statistical parameters can then be interpolated or regressed in order to derive boundaries for speed segments where no training data is available. A comparison of the use of a statistical decision boundary and support vector boundaries, whose inputs have been centralized and whitened with these statistical parameters, will reveal a stronger classification approach. These methods were validated on data gathered from an experimental gearbox and motor apparatus operating at variable speeds; the results indicate a high degree of separability between data from healthy and faulted states—providing exceptional classification error.