DOCTORAL CANDIDATE
Songhao Gao

SUPERVISORS
Prof. Jerome Antoni, Institut National des Sciences Appliquées de Lyon (INSA-L)
Prof. Zhongxiao Peng, University of New South Wales (UNSW)

Mechanical, Industrial and Aerospace Engineering – mechanical system dynamics and modelling, Measurements and Signal Processing, and Machine Condition Monitoring

Condition monitoring critically relies on signal processing for transforming the raw data (vibration, angular speed, strain, etc.) into interpretable features (scalar indicators, spectra, histogram, etc.). One everlasting challenge is to select the signal processing algorithms among a huge number of candidates, while the best choice is obviously case-dependent. Another challenge is to properly use signal processing algorithms, while they often rely on several critical hyperparameters whose optimal setting is again data dependent. The aim of this research project is to propose a solution to these issues, by making signal processing transparent to the user. It consists in developing a machine learning approach, where each algorithm together with its set of hyperparameters is seen as a probabilistic object in a Bayesian hierarchical framework. The main objective is to simultaneously test several candidate algorithms and select the best ones, or a combination of them, according to a given dataset, and to jointly update the values of the hyperparameters that most likely explain the data.