Filtragem inversa não-linear para estimação de sinais em calorímetros operando a alta taxa de eventos

Abstract

Currently, several areas deal with parameter estimation problems as a crucial step. Often, the solution of such problems is challenged by the presence of noise in the observed data. For cases where the noise is Gaussian, several linear techniques are extensively used, however, when Gaussianity properties are lost, a proper noise description is required. In high-energy experiments, the calorimeter systems are responsable for absorbing and sampling the energy from particled produced by collisions. In the LHC, with the increase of its luminosity level, the signal pile-up effect can be observed in the readout channels from the calorimeter systems in the ATLAS experiment. Such effect introduces a nonlinear component to the noise which was previously modeled by a Guassian distribution, degrading the efficieny of typical energy estimation methods. Therefore, in this work, a nonlinear method based on the Maximum Likelihooh Estimatior (MLE) that uses a Lognormal model for the noise is presented. The performance from the proposed method, called MLE Lognormal, is compared with thre other linear methods: OF2, which is currently used in the ATLAS main hadronic calorimeter (TileCal), COF and MLE Gaussian, which assumes a Gaussian noise for the MLE. Additionally, an analysis on the statistical dependence of the noise random variables is presented, using the ICA technique as a pre-processing in order to cope with the statistical dependency. Several channel occupancy conditions and luminosity were considered where it was possible to observe that the Lognormal distribution presents a better fitting when compared to the Gaussian approach. The performance achieved by the methods are consistent with such observations. Improvements of 3.14%, 28.17% and 3.23% were observed for the MLE Gaussian, OF2 and COF, in the simulated data, and 5.39%, 26.59% and 8.34%, in the real data, respectively. It was also possible to assess the statistical dependence between the noise time samples, where the ICA can be a promising alternative to decrease the mutual information between the noise samples. Reductions from 13.26% to 7.00% in the simulated data and from 18.13% to 4.44% in the real data were observed in this parameter.