Métodos de detecção de outliers para o monitoramento ambiental de espaços urbanos inteligentes via análise multivariada e multidimensional

Abstract

Since 2007, for the first time in human history, more people live in cities than in the countryside. According to projections by the United Nations (UN), the world population is expected to be about 70% urban by 2050. This exponential urban growth brings with it critical and typical problems in cities, such as urban mobility, health, public safety and security. environment pollution. Given that cities are a vast and heterogeneous repository of potential data, the complexity of which is proportional to their size and population, a possible solution to address these problems comes from monitoring events associated with urban data. However, one of the challenges when dealing with this data is to distinguish it as "in the pattern"and "out of the pattern"(outlier) which, in the final analysis, can help or hinder the decision making, for example, of a manager public in a city. This thesis investigates the detection of outliers in smart city applications under three different approaches: offline multivariate approach, where we model the data as matrices, suppressing one of its dimensions, and perform a multivariate analysis using the exploratory factor analysis (EFA) multivariate technique; offline multidimensional approach, where we model the data as a third order tensor, and perform an offline multidimensional analysis using the multidimensional technique Higher-Order Singular Value Decomposition (HOSVD); and online multidimensional approach, where we model the data as a third order tensor and perform an online multidimensional analysis combining the multidimensional HOSVD technique with the sliding window strategy. In order to carry out this research, real data were collected from the Smart Citizen urban environmental monitoring platform, configuring themselves as multidimensional data sets given their dimensions: temporal (moments of occurrence of events), environmental variables (physical measures collected by the sensors) and spatial (analyzed cities). The results obtained revealed for the offline multivariate approach which factors were most influential in the patterns of detected outliers (with an accuracy of 75%), whereas for the multidimensional offline approach an outlier detection model was generated (with an accuracy of 91% ), and finally, for the online multidimensional approach, instantaneous variations of the occurrence of specific events were extracted, efficiently identifying the dynamics of the process (with an accuracy of 95%).