Nonlinear system identification
System identification is a method of identifying or measuring the mathematical model of a system from measurements of the system inputs and outputs. The applications of system identification include any system where the inputs and outputs can be measured and include industrial processes, control systems, economic data, biology and the life sciences, medicine, social systems and many more.
Nonlinear systems are usually defined as any system that is not linear, that is any system that does not satisfy the superposition principle. This contrarian description is often necessary because there are so many types of nonlinear systems. Historically system identification for nonlinear systems, has developed by focusing on specific classes of system and can be broadly categorised into four basic approaches which are defined by a model class (i) Volterra series models, (ii) block structured models, (iii) neural network models, and (iv) NARMAX models.
Volterra series methods
The early work was dominated by methods based on the Volterra series, which in the discrete time case can be expressed as
where u(k), y(k); k = 1, 2, 3, … are the measured input and output respectively and is the lth-order Volterra kernel, or lth-order nonlinear impulse response. The Volterra series is an extension of the linear convolution integral. Most of the earlier identification algorithms assumed that just the first two, linear and quadratic, Volterra kernels are present and used special inputs such as Gaussian white noise and correlation methods to identify the two Volterra kernels. In most of these methods the input has to be Gaussian and white which is a severe restriction for many real processes. These results were later extended to include the first three Volterra kernels, to allow different inputs, and other related developments including the Wiener series. A very important body of work was developed by Wiener, Lee, Bose and colleagues at MIT from the 1940s to the 1960s including the famous Lee and Schetzen method,. While these methods are still actively studied today there are several basic restrictions. These include the necessity of knowing the number of Volterra series terms a priori, the use of special inputs, and the large number of estimates that have to be identified. For example for a system where the first order Volterra kernel is described by say 30 samples, 30x30 points will be required for the second order kernel, 30x30x30 for the third order and so on and hence the amount of data required to provide good estimates becomes excessively large. These numbers can be reduced by exploiting certain symmetries but the requirements are still excessive irrespective of what algorithm is used for the identification.
Because of the problems of identifying Volterra models other model forms were investigated as a basis for system identification for nonlinear systems. Various forms of block structured nonlinear models were introduced or re-introduced at this time,. The Hammerstein model consists of a static single valued nonlinear element followed by a linear dynamic element. The Wiener model is the reverse of this combination so that the linear element is before the static nonlinear characteristic. The Wiener-Hammerstein model consists of a static linear element sandwiched between two dynamic systems, and several other model forms are available. All these models can be represented by a Volterra series but in this case the Volterra kernels take on a special from in each case. Identification consists of correlation based and parameter estimation methods. The correlation methods exploited certain properties of these systems which meant that if specific inputs were used, often white Gaussian noise, the individual elements could be identified one at a time. This resulted in manageable requirements of data and the individual blocks could sometimes be related to components in the system under study. More recent results are based on parameter estimation and neural network based solutions. Many results were introduced and these systems continue to be studied in depth. One problem is that these methods are only applicable to a very special form of model in each case and usually this model form has to be known prior to identification.
Artificial neural networks try loosely to imitate the network of neurons in the brain where computation takes place through a large number of simple processing elements. A typical neural network consists of a number of simple processing units interconnected to form a complex network. Layers of such units are arranged so that data is entered at the input layer and passes through either one or several intermediate layers before reaching the output layer. In supervised learning the network is trained by operating on the difference between the actual output and the desired output of the network, the prediction error, to change the connection strengths between the nodes. By iterating the weights are modified until the output error reaches an acceptable level. This process is called machine learning because the network adjusts the weights so that the output pattern is reproduced. Neural networks have been extensively studied and there are many excellent textbooks devoted to this topic in general, and more focussed textbooks which emphasise control and systems applications,. There are two main problem types that can be studied using neural networks static problems, and dynamic problems. Static problems include pattern recognition, classification, and approximation. Dynamic problems involve lagged variables and are more appropriate for system identification and related applications. Depending on the architecture of the network the training problem can be either nonlinear-in-the-parameters which involves optimisation or linear-in-the-parameters which can be solved using classical approaches. The training algorithms can be categorised into supervised, unsupervised, or reinforcement learning. Neural networks have excellent approximation properties but these are usually based on standard function approximation results using for example the Weierstrass Theorem that applies equally well to polynomials, rational functions, and other well-known models. Neural networks have been applied extensively to system identification problems which involve nonlinear and dynamic relationships. However, classical neural networks are purely gross static approximating machines. There is no dynamics within the network. Hence when fitting dynamic models all the dynamics arise by allocating lagged inputs and outputs to the input layer of the network. The training procedure then produces the best static approximation that relates the lagged variables assigned to the input nodes to the output. There are more complex network architectures, including recurrent networks, that produce dynamics by introducing increasing orders of lagged variables to the input nodes. But in these cases it is very easy to over specify the lags and this can lead to over fitting and poor generalisation properties. Neural networks have several advantages; they are conceptually simple, easy to train and to use, have excellent approximation properties, the concept of local and parallel processing is important and this provides integrity and fault tolerant behaviour. The biggest criticism of the classical neural network models is that the models produced are completely opaque and usually cannot be written down or analysed. It is therefore very difficult to know what is causing what, to analyse the model, or to compute dynamic characteristics from the model. Some of these points will not be relevant to all applications but they are for dynamic modelling.
The nonlinear autoregressive moving average model with exogenous inputs (NARMAX model) can represent a wide class of nonlinear systems, and is defined as
where y(k), u(k) and e(k) are the system output, input, and noise sequences respectively; , , and are the maximum lags for the system output, input and noise; F[•] is some nonlinear function, d is a time delay typically set to d = 1.The model is essentially an expansion of past inputs, outputs and noise terms. Because the noise is modelled explicitly unbiased estimates of the system model can be obtained in the presence of unobserved highly correlated and nonlinear noise. The Volterra, the block structured models and many neural network architectures can all be considered as subsets of the NARMAX model. Since NARMAX was introduced, by proving what class of nonlinear systems can be represented by this model, many results and algorithms have been derived based around this description. Most of the early work was based on polynomial expansions of the NARMAX model. These are still the most popular methods today but other more complex forms based on wavelets and other expansions have been introduced to represent severely nonlinear and highly complex nonlinear systems. A significant proportion of nonlinear systems can be represented by a NARMAX model including systems with exotic behaviours such as chaos, bifurcations, and subharmonics. While NARMAX started as the name of a model it has now developed into a philosophy of nonlinear system identification,. The NARMAX approach consists of several steps:
- Structure detection: which terms are in the model
- Parameter estimation: determine the model coefficients
- Model validation: is the model unbiased and correct
- Prediction: what is the output at some future time
- Analysis: what are the dynamical properties of the system
Structure detection forms the most fundamental part of NARMAX. For example a NARMAX model which consists of one lagged input and one lagged output term, three lagged noise terms, expanded as a cubic polynomial would consist of fifty six possible candidate terms. This number of candidate terms arises because the expansion by definition includes all possible combinations within the cubic expansion. Naively proceeding to estimate a model which includes all these terms and then pruning will cause numerical and computational problems and should always be avoided. However, only a few terms are often important in the model. Structure detection, which aims to select terms one at a time, is therefore critically important. These objectives can easily be achieved by using the Orthogonal Least Squares  algorithm and its derivatives to select the NARMAX model terms one at a time. These ideas can also be adapted for pattern recognition and feature selection and provide an alternative to principle component analysis but with the advantage that the features are revealed as basis functions that are easily related back to the original problem.
NARMAX methods are designed to do far more than to just find the best approximating model. System identification can be divided into two aims. The first involves approximation where the key aim is to develop a model that approximates the data set such that good predictions can be made. There are many applications where this approach is appropriate, for example in time series prediction of the weather, stock prices, speech, target tracking, pattern classification etc. In such applications the form of the model is not that important. The objective is to find an approximation scheme which produces the minimum prediction errors. A second objective of system identification, which includes the first objective as a subset, involves much more than just finding a model to achieve the best mean squared errors. This second aim is why the NARMAX philosophy was developed and is linked to the idea of finding the simplest model structure. The aim here is to develop models that reproduce the dynamic characteristics of the underlying system, to find the simplest possible model, and if possible to relate this to components and behaviours of the system under study. The core aim of this second approach to identification is therefore to identify and reveal the rule that represents the system. These objectives are relevant to model simulation and control systems design, but increasingly to applications in medicine, neuro science, and the life sciences. Here the aim is to identify models, often nonlinear, that can be used to understand the basic mechanisms of how these systems operate and behave so that we can manipulate and utilise these. NARMAX methods have also been developed in the frequency and spatio-temporal domains.
- Nelles O. "Nonlinear System Identification: From Classical Approaches to Neural Networks". Springer Verlag,2001
- Billings S.A. "Nonlinear System Identification: NARMAX Methods in the Time, Frequency, and Spatio-Temporal Domains". Wiley, 2013
- Schetzen M. "The Volterra and Wiener Theories of Nonlinear Systems". Wiley, 1980
- Rugh W.J. "Nonlinear System Theory – The Volterra Wiener Approach". Johns Hopkins University Press,1981
- Billings S.A. "Identification of Nonlinear Systems: A Survey". IEE Proceedings Part D 127(6), 272–285,1980
- Haber R., Keviczky L "Nonlinear System Identification-Input Output Modeling Approach". Vols I & II, Kluwer,1980
- Haykin S. "Neural Networks: A Comprehensive Foundation". McMillan,1999
- Warwick K, Irwin G.W., Hunt K.J. "Neural Networks for Control and Systems". Peter Peregrinus, 1992
- Lennart Ljung: System Identification — Theory For the User, 2nd ed, PTR Prentice Hall, Upper Saddle River, N. J., 1999.
- R. Pintelon, J. Schoukens, System Identification: A Frequency Domain Approach, IEEE Press, New York, 2001. ISBN 978-0-7803-6000-6
- T. Söderström, P. Stoica, System Identification, Prentice Hall, Upper Saddle River, N.J., 1989. ISBN 0-13-881236-5
- R. K. Pearson: Discrete-Time Dynamic Models. Oxford University Press, 1999.
- P. Marmarelis, V. Marmarelis, V. Analysis of Physiological Systems, Plenum, 1978.
- K. Worden, G. R. Tomlinson,Nonlinearity in Structural Dynamics, Institute of Physics Publishing, 2001.