Dynamic Characterization of non-Sationary Signals Using Entropy-based Relevance Analysis

Nowadays deal with data of high dimension, that contain complex signals whose information can be contaminated noise, redundancy or irrelevant information can be a recurring problem in processes such as EEG signals or signs of vibrations produced by machines rotating basis. In order to solve this pro...

Full description

Autores:: López Montes, Juan Camilo

Tipo de recurso:

Fecha de publicación:: 2019

Institución:: Universidad Nacional de Colombia

Repositorio:: Universidad Nacional de Colombia

Idioma:: spa

Description
Summary:	Nowadays deal with data of high dimension, that contain complex signals whose information can be contaminated noise, redundancy or irrelevant information can be a recurring problem in processes such as EEG signals or signs of vibrations produced by machines rotating basis. In order to solve this problem, you can deploy an analysis phase of relevance to select the most important components and reduce the high computational cost. However, there is still no standard method for selecting relevant components that may vary according to the object of study, therefore, in order to develop a precise methodology, inherent information should be included and an understanding of the problem. This thesis proposes three methodologies to reveal relevant patterns in the dimensions of space, time and frequency considering the scenario of the experiment. The proposal consists of two stages the first is the feature extraction stage and the second stage is where the characteristics guided by the labels are compared, the methodology which was tested in the vibration and EEG signals, focused on the Motor imagery task (MI). In the MI task the methodology is used to improve the performance of classification in tasks of Brain Computer Interface (BCI), the procedure is to filter each channel in different bands to extract a set of characteristics based on the entropy, subsequently, the characteristics of class one against class two are compared by a measure of dissimilarity, in order to select the set of components that best discriminates the classes, finally, the significant components will be smaller together that achieves the best classification success. The proposal consists of two stages of which can be adjusted according to the task, first the method of extraction of characteristics which must be selected by the type of signal, the type of extraction selected must be able to codify the relevant dynamics of the signal, second the measure of similarity is selected according to the type of features and its dimension. In this document we tested two different stage configurations in order to show the versatility of our method and reveal which configuration better encodes relevant dynamics. The first configuration is based on characterizing the bands filtered by the entropy of Renyi and Ordinal symbolic Dynamics to perform a statistical test guided by the classes on these revealing the level of significance of each band. The second configuration characterized by building a kernel based on the cross-correlation of each band, comparing the array built against the kernel of labels per CKA getting the value relevance of each band and channel. For MI task our methodology was tested in three different BCI datasets and compared against different methods of the state of the art where our relevance analysis statistically improves the classification of MI tasks with a lesser amount of channels and frequency bands. For vibration signals, the condition of the machine must be evaluated by a health index reveals the current state of the rotary machine and identifies early failures in it, first encodes each sensor record by Ordinal Symbolic Dynamical (OSD ) and is built a probability density function (FDP) based on the coding, to be compared against sound records by the Jensen divergence. The order of the OSD must be adjusted optimally because it controls the number of symbols that can exist in the dictionary and therefore the quantity of bins or events of the FDP. This parameter was adjusted according to the state of the art and taking in as far as the order that is directly proportional to the computational cost. In comparison with state methods art our proposed approach succeeds in identifying the start of degradation of bearing more early.In general, the relevance analysis allows a considerable reduction of characteristics, this facilitates the physiological interpretation of the experiments and can improve the performance and the computational cost of the systems that use these characteristics

Dynamic Characterization of non-Sationary Signals Using Entropy-based Relevance Analysis

Publicaciones similares