Optimization of sequential convolutional networks based on improved hunter-prey algorithm remaining engine life prediction

Abstract

The Remaining Life Prediction of the aeroengine is an important link to realize health detection in maintenance based on condition. To improve the prediction accuracy of the Remaining Useful Life (RUL) of the aero engine, the grid search method is adopted to optimize the hyperparameters in the traditional data-driven life prediction algorithm. In the process, it is easy to fall into problems such as local optimization and high dimensional redundancy characteristics with long life data of aero engine and strong sequence. An improved Hunter Prey Optimizer (HPO) is proposed to optimize the Temporal Convolutional Network (TCN) model THPO-TCN. By constructing causal convolution to capture the high-order timing features of fused multi-sensor data, and using expansive convolution to ensure the capture of the medium and long-term dependencies of the time series, the quality of the initial solution of the global optimization algorithm HPO is guaranteed by using chaotic initialization method (Tent). Then the HPO algorithm is improved to find the optimal hyperparameters of the TCN temporal network by introducing refined reverse population and Gaussian variation strategies. The results of aero engine degradation simulation data show that the proposed THPO-TCN network structure model has a significant advantage over TCN, LSTM, GA-TCN, and the unimproved HPO-TCN network model in the prediction accuracy of engine residual life.