A health monitoring method based on multivariate-time series adaptive gated recurrent unit transfer learning model for coal mill system

The economic loss caused by shutdown and fault of coal mill is huge. By effective prognostics and health management (PHM) for health monitoring and fault detection of coal mill system, the overall maintenance cost of coal-fired power plant can be minimized. Therefore, a health monitoring method based on multivariate-time series adaptive gated recurrent unit transfer learning model is proposed. Firstly, LightGBM and correlation analysis are employed to screen the feature variables. Secondly, a multivariate-time series adaptive gated recurrent unit (MTS-AdaGRU) is developed to construct a normal behavior model of coal mill system. In this model, the temporal distribution characterization is used to divide the original sequences into K periods with the least similar distribution. The factorized temporal mixing strategy is adopted to extract the time dependence of K periods, respectively. The common feature of different periods is learned by the temporal distribution matching. Thirdly, a health degree based on Jensen–Rényi divergence is proposed to implement the health assessment, which is carried out by calculating the difference between the actual value and model output value. The effectiveness of the proposed method in health monitoring of coal mill system is verified on the collected actual operation data of coal mill system.