Real Time Model Adaptation for Non-Stationary Systems

This proposed research describes a new approach of modelling for the non-stationary data sets, which are commonly generated in many systems including Radar, Sonar, communications, instrumentation, seismic exploration, speech processing and recognition, etc. This proposal falls into the general area of non-stationary processing and particularly addresses the challenges that signals are to be extracted from non-stationary, spatially and temporally correlated noise. It also relates to the issue of low size, weight and power . Signal processing functions usually perform based on a pre-set model, or the system structure is fixed. Although this provides a simple solution, it is highly inefficient especially for non-stationary systems that are common in practice. This makes it highly desirable to adapt the model so that it can capture the true underlying dynamics and predicts accurately the output for unseen data. This proposed research would particularly focus on on-line real time model adaptation approaches which are an important class of model construction algorithms that deal with model structure and/or parameter updating on the arrival of new data. Alternatively, flexible design with model adaptation for non-stationary systems is potentially possible in practical realisations due to recent technological advances where a large portion of the hardware processing is implemented in software. Therefore, when the maximum potential capability of a system is higher than the requirement, spare resource can be saved for other system functions. This is particularly useful for many modern systems in which limited computing resource is usually shared by multiple functions, which would eventually enable the use hardware of reduced size, weight and power (SWAP).Against this background, however, there is a lack of generic tools/methodologies to deal with the problem as on how to perform the model structural changes as demanded by the processes. Some nonlinear model structure identification algorithms are too slow for real-time applications. Most current real-time algorithms, on the other hand, are ad hoc rather than principled approaches. Hence the resultant model quality is not optimal from a statistical point of view.In this programme, we propose to develop a hybrid, flexible yet principled approach for optimum on-line adaptive modelling by means of minimal model structure determination and simultaneous parameter estimation. The aim of the proposal is to introduce a new technique for the adaptive modelling of complex nonlinear dynamical systems in real time and noisy environments. The proposed algorithms would be validated based on mathematical derivations, proofs and extensive simulations in comparison with competitive algorithms using numerical examples on simulated and realistic benchmark data sets.