Double regulation Levenberg–Marquardt neural networks: an aero-engine fuel flow prediction method

Fuel flow of aero-engine is a very important gas path performance monitoring parameter. It is a highly nonlinear time series data with a large amount of noise. Accurate prediction of fuel flow is a great challenge. Levenberg–Marquardt (LM) neural networks have become a popular method of time series prediction due to its high optimization efficiency. However, in the process of network training, the highly nonlinear data make the LM neural networks generate a large truncation error when calculating the falling direction of loss. This error makes the loss large and cannot be effectively reduced and even continues to increase the loss to cause iterative divergence. In order to solve this problem, a LM neural network with improved optimization strategy, namely reconstructed double regulation LM neural network, is proposed in this paper. First, the aero-engine fuel flow sequence is decomposed based on the phase space reconstruction technology to reduce the nonlinearity of the original data to improve the prediction accuracy of the neural networks model. Secondly, a double regulation optimization algorithm for network weights is proposed, which combines the method of adaptively changing the gradient descent direction and the convergence step size to prevent the neural networks from easily falling into local minimum. The reconstructed double regulation LM neural network is used to predict the fuel flow of a certain type of engine. It can be seen from the simulation results that this improved neural network has the highest prediction accuracy and can reduce the mean absolute error and mean absolute percentage error of engine fuel flow prediction to 36.41 pounds per hour and 1.18, respectively. At the same time, the double regulation optimization algorithm can greatly improve the error convergence speed and save the calculation time according to the trend chart of the neural networks prediction error.