Artificial Neural Network Approaches for Characterizing the Frequency Content of Signals

Slides

The slides from this talk are available here.

 

Title

Artificial Neural Network Approaches for Characterizing the Frequency Content of Signals

 

Abstract

The utility of Artificial Neural Networks (ANNs) lies in the fact that they can be used to infer a function from observations. Indeed, these non-parametric models are able to learn from and to adapt to their environment with weak assumptions. They are appropriate to handle non-linearities and non-stationarities if used with a good understanding of the underlying theory of the task to be solved or of the signal to be analyzed.

 

Our approach consists in formalizing the signals by analytical expressions, to design appropriate inputs and then to use a supervised learning where the reference comes from a measured signal. Then, learning with the neural approach allows estimating the expression and recovering some physically characteristics of the task or of the signal.

 

This is illustrated with Adaline (Adaptive Linear Element) and linear Multi-Layer Perceptron (MLP) networks with applications in processing biomedical and from power system signals. Original neural techniques are therefore proposed for estimating Fourier series representative of any periodic signal. As a result, the amplitudes of the harmonic terms present in the measured signal can be individually identified. Neural techniques are also employed to fit recursive expressions of signals and tasks such as instantaneous frequency tracking, angular or shift phase estimation, and fault detection and diagnosis can be achieved in real-time. Once trained, some interpretable parameters can be obtained directly from the weights which can be considered as physically interpretable. Relevant study cases are provided and performances are evaluated. Simulation and experimental results show the effectiveness of these approaches.

 

About the Presenter

Pr. Patrice WIRA

University of Haute Alsace

MIPS Laboratory (Modelization, Intelligence, Processes and Systems)

4 rue des Frères Lumière, 68093 Mulhouse, France

Tel. : (+33) 3 89 33 60 82       -  Fax  : (+33) 3 89 33 60 84

Email : patrice.wira@uha.fr    -  Web : http://www.trop.uha.fr/ 

 

Patrice Wira is a Full Professor with the MIPS Laboratory (Modelization, Intelligence, Processes and Systems) at the University of Haute Alsace, Mulhouse, France. He is the author or coauthor of more than 100 publications in the field of signal processing, adaptive signal processing, learning systems, artificial neural networks, neural control approaches, neural network implementations, and intelligent techniques. He has contributed new ideas in design and synthesis of neural approaches. Particularly, he set neural networks to control robot arms with a high number of degrees of freedom in visual servoing tasks. He also developed neural network approaches to identify harmonic characteristics for compensating tasks and to detect failures and diseases from signals issued from electrical power systems and biomedical applications.