An Approach to Outlier Detection and Smoothing Applied to a Trajectography Radar Data

Abstract

The tracking of aerospace engines is reasonably achieved through a trajectography radar system that generally yields a disperse cloud of samples on tridimensional space, which roughly describes the engine trajectory. It is proposed an approach on cleaning radar data to yield a wellbehaved and smooth output curve that could be used as basis for instant and further analysis by radar specialists. This approach consists on outlier detection and smoothing phases based on established techniques such as Hampel filter and local regression (LOESS). To prove the effectiveness of the approach, both filtered and unfiltered data are submitted to an extrapolation method, and the results are compared.

Comparison Between Lotka-Volterra and Multivariate Autoregressive Models of Ecological Interaction Systems

“…One of the simplest approaches is the three-point method, where the slope at time point t k is taken as the average of the slopes at time points t k-1 and t k+1 (Burden, Faires, & Burden, 1993;Voit & Almeida, 2003). More sophisticated methods were reviewed in (Cleveland & Grosse, 1991;Eilers & Marx, 1996;Batista Júnior & Pires, 2014). For long, dense time series, moving average and collocation methods with or without roughness penalty (Ramsay et al, 2007) are often very effective.…”

“…Expressed differently, if the noise is left unchecked, its effect on the estimated values of the slopes tends be higher than its effect on the values of the variables. We explored a number of methods for smoothing the time course data and keeping the noise in check (Eilers, 2003;Vilela et al, 2007;Batista Júnior & Pires, 2014), cognizant of the fact that empirical raw data alone do not provide enough information of what is noise and what is relevant signal in the dynamics of the phenomenon under study. One of the simplest approaches is the three-point method, where the slope at time point t k is taken as the average of the slopes at time points t k-1 and t k+1 (Burden, Faires, & Burden, 1993;Voit & Almeida, 2003).…”

Balancing Complex Signals for Robust Predictive Modeling

“…Computer scientists consider the outliers as complex signals [2]. Although such signals may have an adverse impact on the performance of model, their removal is not always legitimate [12] as they may carry important information. There is no consensus on a solid mathematical definition for outliers; however, some statistical tests are available for finding candidate outliers [12].…”

“…Although such signals may have an adverse impact on the performance of model, their removal is not always legitimate [12] as they may carry important information. There is no consensus on a solid mathematical definition for outliers; however, some statistical tests are available for finding candidate outliers [12]. The main contribution of this paper is to devise a novel technique, Complex Signal Balancing (CSB), for training models with outliers until an acceptable performance for a given dataset is achieved.…”

“…Outlier values notably vary from the data distribution. Although, outliers may have an adverse impact on the performance of models, they may also contain important information, and hence, their removal is not always suggested [12]. On the other hand, in modern machine learning, outliers are regarded as complex signals and are not removed from the training datasets.…”