Radar-based respiration monitoring has been increasingly popular among researchers in biomedical fields during the last decades since it is a contactless monitoring technique. It is very convenient for subjects because it does not impose any restrictions on subjects or require their cooperation. Meanwhile, recognizing alternations in respiratory patterns is an important early clue of the diagnosis of several cardiorespiratory diseases. Thus, a study of biomedical radar-based respiration monitoring and respiratory pattern classification is carried out in this thesis. Radar-based respiration monitoring technology has a shortcoming that the collected respiratory signal will be easily distorted by the body movement of the monitoring subjects or disturbed by environment noise because of the contactless measurement attribute. This shortcoming limits the application of the respiratory pattern classification model, that is, the existing models cannot be applied automatically since the distorted respiratory signal needs to be manually filtered out ahead of the classification. In this study, a new respiratory pattern classification strategy, which can be implemented full-automatic, is proposed. In this strategy, a class “moving” is introduced to classify the distorted signal, and the sampling window length is shortened to reduce the effect caused by the signal distortion. A performance requirement for the continuous respiratory pattern classification is also proposed based on its expected function that can alert the occurrence of the abnormal breathing patterns. Several models which can meet the proposed performance requirement are developed in this thesis based on the state-of-the-art pattern classification technique and the time-series-based shapelet transform algorithm. The proposed models can classify four breathing patterns including eupnea, Cheyne Stokes respiration, Kussmaul breathing and apnea. A radar-collected respiratory signal database is built in this study, and a respiration simulation model which can generate breath samples for pattern classification is developed in this thesis. The proposed models were tested and validated in batch and stream processing manner with independently collected data and continuously collected data, respectively.

Nos travaux décrits dans cette thèse portent sur l’apprentissage d’une représentation pour la classification automatique basée sur la découverte de motifs à partir de séries temporelles. L’information pertinente contenue dans une série temporelle peut être encodée temporellement sous forme de tendances, de formes ou de sous-séquences contenant habituellement des distorsions. Des approches ont été développées pour résoudre ces problèmes souvent au prix d’une importante complexité calculatoire. Parmi ces techniques nous pouvons citer les mesures de distance et les représentations de l’information contenue dans les séries temporelles. Nous nous concentrons sur la représentation de l’information contenue dans les séries temporelles. Nous proposons un cadre (framework) pour générer une nouvelle représentation de séries temporelles basée sur la découverte automatique d’ensembles discriminants de sous-séquences. Cette représentation est adaptée à l’utilisation d’algorithmes de classification classiques basés sur des attributs. Le framework proposé transforme un ensemble de séries temporelles en un espace d’attributs (feature space) à partir de sous-séquences énumérées des séries temporelles, de mesures de distance et de fonctions d’agrégation. Un cas particulier de ce framework est la méthode notoire des « shapelets ». L’inconvénient potentiel d’une telle approache est le nombre très important de sous-séquences à énumérer en ce qu’il induit un très grand feature space, accompagné d’une très grande complexité calculatoire. Nous montrons que la plupart des sous-séquences présentes dans un jeu de données composé de séries temporelles sont redondantes. De ce fait, un sous-échantillonnage aléatoire peut être utilisé pour générer un petit sous-ensemble de sous-séquences parmi l’ensemble exhaustif, en préservant l’information nécessaire pour la classification et tout en produisant un feature space de taille compatible avec l’utilisation d’algorithmes d’apprentissage automatique de l’état de l’art avec des temps de calculs raisonnable. On démontre également que le nombre de sous-séquences à tirer n’est pas lié avec le nombre de séries temporelles présent dans l’ensemble d’apprentissage, ce qui garantit le passage à l’échelle de notre approche. La combinaison de cette découverte dans le contexte de notre framework nous permet de profiter de techniques avancées (telles que des méthodes de sélection d’attributs multivariées) pour découvrir une représentation de séries temporelles plus riche, en prenant par exemple en considération les relations entre sous-séquences. Ces résultats théoriques ont été largement testés expérimentalement sur une centaine de jeux de données classiques de la littérature, composés de séries temporelles univariées et multivariées. De plus, nos recherches s’inscrivant dans le cadre d’une convention de recherche industrielle (CIFRE) avec Arcelormittal, nos travaux ont été appliqués à la détection de produits d’acier défectueux à partir des mesures effectuées par les capteurs sur des lignes de production. / Our research described in this thesis is about the learning of a motif-based representation from time series to perform automatic classification. Meaningful information in time series can be encoded across time through trends, shapes or subsequences usually with distortions. Approaches have been developed to overcome these issues often paying the price of high computational complexity. Among these techniques, it is worth pointing out distance measures and time series representations. We focus on the representation of the information contained in the time series. We propose a framework to generate a new time series representation to perform classical feature-based classification based on the discovery of discriminant sets of time series subsequences (motifs). This framework proposes to transform a set of time series into a feature space, using subsequences enumerated from the time series, distance measures and aggregation functions. One particular instance of this framework is the well-known shapelet approach. The potential drawback of such an approach is the large number of subsequences to enumerate, inducing a very large feature space and a very high computational complexity. We show that most subsequences in a time series dataset are redundant. Therefore, a random sampling can be used to generate a very small fraction of the exhaustive set of subsequences, preserving the necessary information for classification and thus generating a much smaller feature space compatible with common machine learning algorithms with tractable computations. We also demonstrate that the number of subsequences to draw is not linked to the number of instances in the training set, which guarantees the scalability of the approach. The combination of the latter in the context of our framework enables us to take advantage of advanced techniques (such as multivariate feature selection techniques) to discover richer motif-based time series representations for classification, for example by taking into account the relationships between the subsequences. These theoretical results have been extensively tested on more than one hundred classical benchmarks of the literature with univariate and multivariate time series. Moreover, since this research has been conducted in the context of an industrial research agreement (CIFRE) with Arcelormittal, our work has been applied to the detection of defective steel products based on production line's sensor measurements.

Uma série temporal é uma sequência ordenada pelo tempo de valores reais. Dado que inúmeros fenômenos do dia-a-dia podem ser representados por séries temporais, há grande interesse na mineração de dados temporais, em especial na tarefa de classificação. Recentemente foi introduzida uma nova primitiva de séries temporais chamada shapelet, que é uma subsequência que permite a classificação de séries temporais de acordo com padrões locais. Na transformada shapelet estas subsequências se tornam atributos em uma matriz de distância que mede a dissimilaridade entre os atributos e as séries temporais. Para obter a transformada é preciso escolher alguns shapelets dos inúmeros possíveis, seja pelo efeito de evitar overfitting ou pelo fato de que é computacionalmente caro obter todos. Sendo assim, foram elaboradas medidas de qualidade para os shapelets. Tradicionalmente tem se utilizado a medida de ganho de informação, porém recentemente foi proposto o uso da f-statistic, e nós propomos neste trabalho uma nova denominada in-class transitions. Em nossos experimentos demonstramos que a inclass transitions costuma obter a melhor acurácia, especialmente quando poucos atributos são utilizados. Além disso, propomos o uso de amostragem aleatória nos shapelets para reduzir o espaço de busca e acelerar o processo de obtenção da transformada. Contrastamos a abordagem de amostragem aleatória contra uma em que só são exploradas shapelets de determinados tamanhos. Nossos experimentos mostraram que a amostragem aleatória é mais rápida e requer a computação de um menor número de shapelets. De fato, obtemos os melhores resultados ao amostrarmos 5% dos shapelets, mas mesmo a uma amostragem de 0,05% não foi possível notar uma degradação significante da acurácia. / A time series is a time ordered sequence of real values. Given that numerous daily phenomena that can be described by time series, there is a great interest on its data mining, specially on the task of classification. Recently it was introduced a new time series primitive called shapelets, that is a subsequence that allows the classification of time series by local patterns. On the shapelet transformation these subsequences turn into attributes in a distance matrix that measures the dissimilarity between these attributes and the time series. To obtain the shapelet transformation it is required to choose some shapelets among all of the possible ones, be it to avoid overfitting or because it is too computationally expensive to obtain everyone. Thus, some shapelet quality measurements were created. Traditionally the information gain has been used as the default measurement, however, recently it was proposed to use the f-statistic instead, and in this work we propose a new one called in-class transitions. On our experiments it is shown that usually the in-class transitions achieves the best accuracy, specially when few attributes are used. Moreover, we propose the use of random sampling of shapelets as a way to reduce the search space and to speed up the process of obtaining the shapelet transformation. We contrast this approach with one that explores only shapelets that have a specific length. Our experiments show that random sampling is faster and requires fewer shapelets to be computed. In fact, we got the best results when we sampled 5% of the shapelets, but even at a rate of 0.05% it was not possible to detect a significant degradation of the accuracy.

Uma série temporal é uma sequência ordenada pelo tempo de valores reais. Dado que inúmeros fenômenos do dia-a-dia podem ser representados por séries temporais, há grande interesse na mineração de dados temporais, em especial na tarefa de classificação. Recentemente foi introduzida uma nova primitiva de séries temporais chamada shapelet, que é uma subsequência que permite a classificação de séries temporais de acordo com padrões locais. Na transformada shapelet estas subsequências se tornam atributos em uma matriz de distância que mede a dissimilaridade entre os atributos e as séries temporais. Para obter a transformada é preciso escolher alguns shapelets dos inúmeros possíveis, seja pelo efeito de evitar overfitting ou pelo fato de que é computacionalmente caro obter todos. Sendo assim, foram elaboradas medidas de qualidade para os shapelets. Tradicionalmente tem se utilizado a medida de ganho de informação, porém recentemente foi proposto o uso da f-statistic, e nós propomos neste trabalho uma nova denominada in-class transitions. Em nossos experimentos demonstramos que a inclass transitions costuma obter a melhor acurácia, especialmente quando poucos atributos são utilizados. Além disso, propomos o uso de amostragem aleatória nos shapelets para reduzir o espaço de busca e acelerar o processo de obtenção da transformada. Contrastamos a abordagem de amostragem aleatória contra uma em que só são exploradas shapelets de determinados tamanhos. Nossos experimentos mostraram que a amostragem aleatória é mais rápida e requer a computação de um menor número de shapelets. De fato, obtemos os melhores resultados ao amostrarmos 5% dos shapelets, mas mesmo a uma amostragem de 0,05% não foi possível notar uma degradação significante da acurácia. / A time series is a time ordered sequence of real values. Given that numerous daily phenomena that can be described by time series, there is a great interest on its data mining, specially on the task of classification. Recently it was introduced a new time series primitive called shapelets, that is a subsequence that allows the classification of time series by local patterns. On the shapelet transformation these subsequences turn into attributes in a distance matrix that measures the dissimilarity between these attributes and the time series. To obtain the shapelet transformation it is required to choose some shapelets among all of the possible ones, be it to avoid overfitting or because it is too computationally expensive to obtain everyone. Thus, some shapelet quality measurements were created. Traditionally the information gain has been used as the default measurement, however, recently it was proposed to use the f-statistic instead, and in this work we propose a new one called in-class transitions. On our experiments it is shown that usually the in-class transitions achieves the best accuracy, specially when few attributes are used. Moreover, we propose the use of random sampling of shapelets as a way to reduce the search space and to speed up the process of obtaining the shapelet transformation. We contrast this approach with one that explores only shapelets that have a specific length. Our experiments show that random sampling is faster and requires fewer shapelets to be computed. In fact, we got the best results when we sampled 5% of the shapelets, but even at a rate of 0.05% it was not possible to detect a significant degradation of the accuracy.