Bayesian Sensor Estimation for Machine Condition Monitoring

We present a Bayesian framework to tackle the problem of sensor estimation, a critical step of fault diagnosis in machine condition monitoring. A Gaussian mixture model is employed to model the normal operating range of the machine. A Gaussian random vector is introduced to model the possible deviations of the observed sensor values from their corresponding normal values. Different levels of deviations are elegantly handled by the covariance matrix of this random vector, which is estimated adaptively for each input observation. Our algorithm doesn't require faulty operation training data, as desired by previous methods. Significant improvements over previous methods are achieved in our tests.