Analysis of space-filling excitation signals and dynamic models for nonlinear system identification of dynamic processes of a Diesel engine

Abstract

The paper investigates the influence of design of experiments (DoE) methods in combination with various mathematical architectures for the goal of an accurate Diesel engine identification. Three step-based excitation signals – amplitude-modulated pseudo random binary signal (APRBS), optimized nonlinear input signal (OMNIPUS), and the stacked global optimized amplitude time signal (sGOATS) – are analyzed regarding their space-filling property, a good excitation of all frequencies, and achievable model quality. Six different model architectures are used: a finite impulse response multilayer perceptron (MLP-NFIR), a long short-term memory (LSTM) network, a gated recurrent unit (GRU) network, a local model network with regularized finite impulse response models (LMN-NRFIR), a local model network with auto-regressive with exogenous inputs models (LMN-NARX), and a local model state space network (LMSSN). The model architectures are designed to be suitable for low-performance microcontrollers. As a process, a real-world system (Diesel engine) with high system complexity due to multiple feedback paths, strong nonlinearities, and infeasible regions is chosen, which allows a proper investigation of the data-based methods. The sGOATS achieves the best space-filling property, an even excitation of all frequencies, and provides good model qualities. It is shown that it is worth to optimize the excitation signals in order to get better models. Regarding the model architectures, the LMN-NARX and LMN-NRFIR perform worst, whereas the LSTM, GRU, and LMSSN slightly surpass the MLP-NFIR.