On a Grey Box Modelling Framework for Nonlinear System Identification
In machine learning, black box models are often used to make excellent predictions of system behaviour. They are especially useful where the physics of a system is unknown or hard to model. This paper examines whether prior knowledge of certain physical properties of a system, encoded in a white box model, can be incorporated into black box methods to improve predictive performance. A combination of genetic algorithm optimisation of the white box model and Gaussian process regression on the residual error is presented as an improved method for system identification. This approach retains physical insight into the behaviour of the system while also reducing the error. Comparisons are made between pure white and black box models and the combined grey box model for several test applications. It is found that the combined model has significant advantages in predictive accuracy. This is especially seen in the case of nonlinear models. Here the full physics of the system is often too complicated or inaccessible to be modelled accurately with a white box method, but also the state space relationships are sufficiently complicated to make black box modelling equally challenging. The use of the proposed grey box method can reduce the complexity of the relationship that the black box is attempting to represent, leading to gains in accuracy. As a bonus, training time is reduced, as less complicated techniques are required to identify the process accurately.
KeywordsNonlinear system identification White box Black box Grey box Gaussian processes
The authors wish to thank Ramboll for providing funding for this work and in particular Ulf Tyge Tygesen whom they are working with directly.
- 2.Noël, J., Kerschen, G.: 10 years of advances in nonlinear system identification in structural dynamics: a review. In: Proceedings of ISMA 2016-International Conference on Noise and Vibration Engineering (2016)Google Scholar
- 6.Bishop, C.: Pattern Recognition and Machine Learning (Information Science and Statistics), 1st edn. 2006, Springer (2007)Google Scholar
- 9.Worden, K.: Is system identification just machine learning? Workshop on Nonlinear System Identification Benchmarks (2016)Google Scholar
- 12.Schoukens, M., Mattsson, P., Wigren, T., Noël, J.P.: Cascaded tanks benchmark combining soft and hard nonlinearities (2016)Google Scholar
- 14.Eberhart, K.J., et al.: A new optimizer using particle swarm theory. In: Proceedings of the Sixth International Symposium on Micro Machine and Human Science, vol. 1, pp. 39–43, New York, NY (1995)Google Scholar
- 15.Malik, R., Rahman, T., Hashim, S., Ngah, R.: New particle swarm optimizer with sigmoid increasing inertia weight. Int. J. Comput. Sci. Secur. 1 (2), 35–44 (2007)Google Scholar
- 16.Sun, J., Feng, B., Xu, W.: Particle swarm optimization with particles having quantum behavior. Congress on Evolutionary Computation (2004)Google Scholar
- 19.Worden, K., Manson, G., Cross, E.: Higher-order frequency response functions from Gaussian process NARX models. In: Proceedings of 25th International Conference on Noise and Vibration Engineering, Leuven (2012)Google Scholar