Sparsity preserving projection aided baselined hyperdisk modeling for interpretable machine health monitoring

Geometric representation optimization models have been widely introduced for machine health monitoring. Nevertheless, their extracted features for machine fault diagnosis lack physical interpretability and health indicators (HIs) have inapparently initial degradation points. In this study, a sparsity preserving projection aided baselined hyperdisk model for interpretable initial fault detection, diagnosis, and degradation assessment is proposed. Herein, a hyperdisk that constructs a proper geometric approximation of a dataset is defined as the cross zone between a hypersphere and an affine hull. Firstly, based on sparsity preserving projection, a physicsinformed methodology is proposed to extract interpretable projection features for immediate fault diagnosis. Simultaneously, based on extracted low-dimensional features, a hyperdisk based degradation modeling methodology is proposed to construct a HI for initial fault detection. Herein, once a baselined hyperdisk model is constructed, a HI can be efficiently computed by simultaneously considering the projection and relative distances of the baselined hyperdisk model. Case studies show that the proposed methodology has superior performances than support vector data description (SVDD), hyperplane based degradation modeling, and other statistical based HIs.

Automated summary

In this study, a sparsity preserving projection aided baselined hyperdisk model is proposed for interpretable initial fault detection, diagnosis, and degradation assessment. The methodology extracts interpretable projection features for immediate fault diagnosis and constructs a HI for initial fault detection based on low-dimensional features. Case studies show that the proposed methodology outperforms SVDD, hyperplane based degradation modeling, and other statistical based HIs.