logo

You are here

Warning message

Attention! This event has already passed.

Resource-Aware Design of Smart Measurement Systems: A Learning-from-examples Approach

Friday, 27 March, 2009 - 15:00
Campus: Brussels Humanities, Sciences & Engineering campus
Anna Marconato
phd defence

Abstract

This thesis deals with the design of accurate and resource-aware learningfrom- examples algorithms (LFEAs) in the context of smart measurement systems. Both discrete and continuous problems are taken into consideration, ranging from various classification tasks to digital sensor compensation and nonlinear system identification. The achieved results are of interest for a very broad set of application areas, such as environmental monitoring, industrial quality control and automatic surveillance.

The motivation for our research has its background in disciplines from heterogeneous fields, including areas such as embedded systems design, optimization methods, evolutionary algorithms, machine learning, measurement theory and system identification theory. The definition and development of design strategies suitable for measurement systems endowed with machine intelligence has allowed to consider these apparently distant areas in a unified framework.

The contribution to the state of the art in the design of intelligent measurement systems is threefold. First we provide a description of LFEA modules in smart measurement systems. To do this, the main aspects of machine learning have to be reconsidered from a metrological point of view. We focus in particular on the analysis of the different uncertainty sources arising in a measurement system, and on the definition of methods used to evaluate such uncertainty.

Secondly, we propose the adoption of a multi-objective optimization paradigm as a key design methodology, so to reduce the computational complexity of the classification or regression function, while keeping the uncertainty level on satisfactory values. This approach is motivated by the unusual resource scarcity conditions in which a machine learning algorithm has to operate. More in details, the dissertation focuses on the study of Support Vector Machines (SVMs), both for classification and regression problems, as a key example of learning-from-examples paradigm. Genetic algorithms are employed in a multi-objective fashion in order to determine the best configuration of the hyperparameters during the model selection phase. Reduced-set methods (i.e. sub-optimal SVM-like techniques) have also been considered, as an alternative approach to control the computational complexity. The latter turns out to be a promising strategy especially when hardware implementation is a fundamental issue.

The third contribution to the state of the art is the application of these resource-aware algorithms to the design of sensor compensation and system identification strategies. As far as the sensor compensation task is concerned, traditional SVM approaches and reduced-set methods have been studied, focusing on the severe constraints imposed by hardware implementation on microcontrollers. For system identification purposes, both linear and nonlinear problems have been addressed. As a first step we have dealt with the problem of designing suitable excitation signals to be given in input to the considered system. Furthermore, we have analyzed the effect on the performance of modifying design choices and parameters (such as the choice of features to be taken into account and the kind of kernel function to be used in the nonlinear case).

The achieved results have been validated via extensive simulations relying on both synthetic and real-world data.

Keywords

measurement systems, uncertainty evaluation, learning-from-examples, support vector machines, multi-objective optimization, genetic algorithms, sensor compensation, system identification.