This paper proposes and analyses a statistical method for uncertainty evaluation of extreme values (minimal or maximal) for measurement results with significantly limited number of observations n = 3…10 and considerable deviation of observation probability density function (PDF) from normal distribution. The method is based on properties of order statistics. It can be used for the uncertainty evaluation of mechanical properties of testing products in a food industry (when minimal values of measurement results are observed) and for the investigation of a number of harmful elements (when maximal values of measurement results are observed).
The uncertainty of measurements associated with the following correction methods: advanced correction of additive linear drift, correction of additive and multiplicative effects, as well as joint correction of a linear drift and systematic additive and multiplicative effects is analyzed in the present article. For each correction method sensitivity coefficients and amplitude responses according to which noise and internal and external interferences influence the corrected measurement result have been determined. Besides uncertainty of reference quantities, the main factors which limit the efficiency of correction are: non-linearity of measurement function including non-linearity of ADC, no idealities of the switching systems and external and internal noises and periodic interferences. The efficiency of correction of systematic additive and multiplicative effects was studied for the multifunction 16 bit PCI DAQ of family NI 6250.
In the paper the new method of the measurement observations processing, based on their comparison (after sorting) with several reference samples, which are corresponded to the models of the general population density distributions (so called reference distributions), is investigated and analyzed. The elements of reference samples are equal to the mathematical expectations of order statistics corresponding to the reference distribution. The mathematical models of the determination of the best result and its standard uncertainty are presented. The effectiveness of proposed method is investigated by the Monte Carlo method for 5 models of general population (Laplace, normal, triangular, uniform and arcsine) with the number of observations equal 9, 19, 29, 39 and 49. The proposed method can be used if the observations number is small. If the observations distribution significantly differs from normal distribution then the proposed method guarantees considerable decreases of the uncertainty result in comparison with the uncertainty of average value.