Integration of a Novel Knowledge-Guided Loss Function With an Architecturally Explainable Network for Machine Degradation Modeling

Abstract

The suitability of a health index (HI) is significant to improve machine diagnostic and prognostic performance. A life cycle HI can be applied to the detection of incipient faults and determination of the first prediction time, and then its monotonic curve after FPT can be used as an immediate variable for prognostic analysis. Although existing data-driven methodologies for HI construction have considerable nonlinear mapping, they lack transparency and interpretability for subsequent studies and analysis. Moreover, these HIs have some unknown burrs and fluctuations. To overcome these limitations, an architecturally explainable network for machine degradation modeling is proposed in this paper. The first three hidden layers are reformulated from three advanced signal processing technologies including Hilbert transform, squared envelope, and Fourier transform to map temporal signals in the time domain into demodulated signals in the squared envelope spectrum domain so as to directly link the proposed network with fault frequencies and their harmonics in the frequency domain. The fourth hidden layer consists of a full connection layer with one hidden node and a novel knowledge-guided loss function specifically designed for exploring and extracting informative degradation features. Case studies are implemented to verify the proposed methodology and show that the proposed HIs have more stable and robust performance compared to state of the art methods. Another superiority of the proposed network lies in its full transparency and interpretability by effectively integrating physics-based signal processing technologies with an interpretable artificial network structure for machine degradation modeling.