Multi-scale signal feature processing for automatic, objective vehicle noise and vibration quality analysis

The paper describes a new signal processing technique for understanding the dynamics of time-varying signals in vehicles: hyperstate analysis. Vehicle noise and vibration are examples of randomly-varying transient or non-stationary signals that are not effectively analyzed with classical spectral analysis techniques. By the use of nested hidden Markov models, hyperstate analysis explicitly identifies transient and nonstationary behavior on many time scales for better signal discrimination. It uses a probability-based framework that allows for automated, objective classification of noisy signals. The technique is applied to engine starting sequences from different types of vehicles. This work demonstrates that hyperstate analysis discerns similarities and differences in randomly-varying signals of this type, and can perform effective automatic, objective classification and signal decomposition for NVH (noise, vibration, and harshness) studies.