Previsão de eventos em ambientes inteligentes com extração de características sequenciais e localização de usuários por modelos ocultos de Markov

• Ano de defesa: 2019
• Autor(a) principal: Igor Pereira Gomes
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA; Programa de Pós-Graduação em Engenharia Elétrica; UFMG
• Palavras-chave em Português: Smart home; Previsão de eventos; Reconhecimento de padrões sequenciais; Extração de características sequenciais; Classificador por arestas de suporte; Grafos de Gabriel; Classificadores sem parâmetros; Aprendizado de métrica; Modelo oculto de Markov; Engenharia elétrica; Automação residencial; Modelo escondido de Markov
• Link de acesso: http://hdl.handle.net/1843/55271

This thesis presents studies done on the Event Prediction problem and the unsupervised User Tracking problem in smart environments, using the data collected by the Minibox home automation system, developed by the Brazilian company Neocontrol. An architecture for solving those problems is proposed and validated, with a sequence of device state changes as its input. The event prediction problem is formulated as a supervised classification problem. A new approach for small and imbalanced sample classification without hyperparameters is proposed, based on Support Edge Classifiers (CHIP-CLAS) and metric learning. Three sequential feature extraction methods were adapted and validated for the event prediction problem, based on the presented smart home data, as a preprocessing step for the classification. For the user tracking problem, a Factorial Hidden Markov Model is employed with its hidden states representing the location for the individuals and its emissions representing the observed device activations. Its parameters are estimated a priori through simple rules based on the floor plan and location for the devices. In the experiments done for the event prediction problem, good results were obtained with the SVM classifier using all three feature extraction methods. With the CLAS classifiers, although results were equivalent to SVMs for a benchmark consisting on 15 different datasets for both CHIP-CLAS and AM-CHIP-CLAS, the observed performance for the event prediction data was far behind, due to small sample size and imbalanced data. Still, the metric learning step proposed for AM-CHIP-CLAS significantly improved the performance comparing to CHIP-CLAS. For the user tracking FHMM model, validation was done through manual inspection with smart home data from the literature and results were consistent with data annotation.