A multiscale stochastic modeling approach to the monitoring of mechanical systems

Presents results on using a statistical model motivated by the wavelet transform to represent non-stationary signals typically encountered in machinery monitoring applications. The authors propose the use of a frame-based system in which the data in each frame is modeled as a multiscale stochastic process. The parameters of a multiscale model are used as features for each frame, where each frame of features is modeled as a sample of a multivariate, multimodal distribution. Classification of machine states based on monitoring signals is performed by comparing likelihood scores for each machine state. The authors present an example of applying the system to data consisting of a superposition of damped sinusoids, as a way of illustrating system performance for the case of transient monitoring signals. They compare their system to one which is trained using a DFT-based (non-time-frequency-based) representation (in particular, LPC coefficients) and show that their system exhibits both superior performance as well as greater robustness to noise in the signals. They also compare results using multiscale parameters versus LPC coefficients for the case of synthesized autoregressive signals and for the case of actual, measured signals from a weld depth monitoring system