Benchmark transformation neural network for health indicator construction under time-varying speed and its application in machinery prognostics

Abstract

A health indicator (HI) is usually applied to predict the remaining useful life (RUL) of machinery. The current HI construction methods typically only focus on constant operating conditions (such as speed or load), rendering them ineffective for variable-speed conditions. To address this gap, this paper proposes a HI construction method for machines under time-varying speed conditions based on benchmark transformation neural networks (BTNN). First, the baseline speed range is determined, and the baseline state observations are identified. Then, with the identified baseline observations, a performance degradation model is established via the double exponential function. Next, BTNN is innovatively proposed using the fitted degradation model, the monitoring state, and speed data to adaptively perform the complex nonlinear mapping from non-baseline observations to baseline values, avoiding the problem of selecting different baseline transformation functions. Compared with the HI constructed by a transformation function and the dimensionless HI, the proposed method unifies state observations at various speeds onto the baseline speed through BTNN, enhancing the comprehensive performance of HI. Comparative experiments on the RUL prediction of turbofan engines and wind turbine bearings reveal that the HI extracted by BTNN exhibits stronger prognosis capabilities than other classical and advanced HIs.