Learning-based design optimization of second-order tracking differentiator with application to missile guidance law

Tracking differentiator is widely used to estimate the differential signal of noisy systems with unknown structures. However, there still lacks an efficient approach to tracking differentiator design and optimization, partly due to the nonlinearity by nature, and the high dimension of design parameters. Thus, this paper proposes a learning-based design optimization method. First, neural networks are trained to approximate the multivariate influence on tracking differentiator performance. The resultant learning-based surrogate model is then used to optimize the parameters of tracking differentiator by differential evolution algorithm. The proposed method is applied to Sigmoid-type tracking differentiator and further compared to the unscented Kalman filter with various input signals. The estimation error of the optimized tracking differentiator is reduced by 65.70% compared to that of the benchmark method. The second-order tracking differentiator is integrated into an angular acceleration-based guidance law. Simulation results of tracking moving targets show that the miss distance and control resource are separately lowered by at least 35.70% and 74.12% as opposed to the crude guidance law.