Contemporary Engineering Sciences, Vol. 11, 2018, no. 86, 4293-4300 HIKARI Ltd, www.m-hikari.com https://doi.org/10.12988/ces.2018.88481 Experimental Analysis of Random Errors for Calibration of Scales Applying Non-Parametric Statistics José Daniel Hernández-Vásquez 1, Cristian Pedraza-Yepes 1, Pedro Grimaldo Pérez 1, Carlos Nuñez Camelo 1 and Jorge González-Coneo 2 1 Program of Mechanical Engineering. Universidad del Atlántico Puerto Colombia - Atlántico, Colombia 2 Department of Energy. Universidad de la Costa Barranquilla-Atlántico, Colombia Copyright 2018 José Daniel Hernández-Vásquez et al. This article is distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Abstract The main objective of the research is to apply the consecrated criteria of nonparametric statistics to evaluate the type of probability distribution that random errors follow in the calibration of scales. This work was motivated, mainly, by the difficulty of accessing relevant and reliable information related to the robust statistical methods applied to metrology, specifically, in the calibration of scales. In this sense, the consolidated results show that the Kolmogorov-Smirnov hypothesis test constitutes an important tool to apply to the field of industrial metrology. Moreover, the statistical analysis of the data confirmed that an adherence of the random errors to a normal probability distribution for a confidence level of 95%. In conclusion, it is confirmed experimentally that it is possible to assume a Gaussian distribution for the errors associated with the calibration of scales and in this way obtain a greater metrological reliability in the different industrial measurement processes. Keywords: Metrology, Random errors, Kolmogorov-Smirnov, Non-parametric statistics
4294 José Daniel Hernández-Vásquez et al. Introduction The statistical analysis of a set of measurements made is an indispensable step in order to characterize a metrological process. In the specialized bibliography there are several studies that use parametric statistical techniques, which characterize the metrological process involved in the measurements made by analog scales [1-2]. Parametric statistics have limitations in terms of the distribution of experimentally obtained data, since, in such an approach, it is necessary to consider the hypothesis that the measurements are modeled by means of a Gaussian distribution. There is a problem related to the measurements made by the analog scales. In these processes, the statistical distribution of the source data cannot be considered a priori as Gaussian or normal [3]. This academic study sought to overcome this problem through the use of nonparametric statistical models (Chi-square and Kolmogorov- Smirnov) [4]. In these situations, an analysis applying Non-Parametric Statistics is useful and extremely necessary. One of the main difficulties in applying this approach is the specification of the statistical distribution of probability to which the experimental data can be associated [5-7]. For this purpose, the literature suggests a series of hypothesis tests that can be applied, for example: Chi-square and Kolmogorov- Smirnov to specify the probability distribution that best models the experimental data. The above with the vision of obtaining the lowest uncertainties associated with the result of the measurement. 2 Normality test: Kolmogorov-Smirnov (K-S) This normality test states that: Null Hypothesis (H0): the sample data follow a Gaussian distribution Alternative Hypothesis (H1): the sample data do not follow a Gaussian distribution In order to confirm that the experimental data indeed follow a normal distribution, the K-S test (at confidence levels of 95% and 99%) was applied. The acceptance criterion for the Null Hypothesis (that automatically rejects the Alternative Hypothesis) is Dmax Dcrit, where Dmax denotes the test statistic and Dcrit a critical value extracted from the Dixon distribution for a given confidence level (in this work, two confidence levels were considered: 95% and 99%). Equation (1) allows calculation of the Dmax value: D max = d n N = F a n e an N (1)
Experimental analysis of random errors for calibration 4295 where the subscript n represents the interval number, F an denotes the cumulative absolute frequency given by equation (2), e an the cumulative nominal frequency given by equation (3) and N the sample size: F an = F an 1 + F in (2) e an = e an 1 + e in (3) In these expressions: F in : absolute frequency of the experimental data; e in : theoretical frequency of the experimental data. The theoretical frequency (e in ) is calculated applying (4): e in = [F(Z 2n ) F(Z 1n )] F in (4) In this expression, F (Z) is associated with the area of the normal (theoretical) distribution; i.e.: F (Z2) and F (Z1) refers, respectively, to the area of the point Z2 to the left tail of a normal distribution and to the area of the point Z1 to the left tail of a normal distribution. Parameters Z1 and Z2 are calculate by equations (5) and (6): Z 1n = x 1 n μ (5) σ Z 2n = x 2 n μ σ In these expressions: x 1n : the lower limit of the data experimental; x 2n : the upper limit of the data experimental; µ: population mean of the data experimental. For the Normal Distribution, the better estimate for the µ is given for x (sample mean), thus, μ x. σ: population standard deviation of the data experimental. For the Normal Distribution, the better estimate for the σ is given for s (sample standard deviation), thus, σ s. 3 Experimental Methodology The experiments were realized at Nutrition Laboratory of Universidad del Atlántico. Mass certificated (E1, F1 and M1) for a Metrology Accredited Laboratory were utilized for totalized 400 experimental points. The instrument used for the experiments corresponding to Analogical Scale Detection (Cap. Max.: 180 kg; Resolution: 0.01 kg). (6)
4296 José Daniel Hernández-Vásquez et al. Figure 1. Analogical Scale Detection (Max. Cap.: 180 kg; Resolution: 0.1 kg). The atmospheric pressure and environment temperature were measured, respectively, with a barometer (U=0.001 mbar/abs; k=2) and thermometer (U=0.1 o C; k=2). Furthermore, the humidity was measured from a hygrometer (U=3 %HR; k=2). These instruments were calibrated in an accredited metrology laboratory [1]. Tables 1 shows the experimental data. Table 1. Calibration of Analogical Scale Detection: Experimental data The atmospheric pressure and environment temperature were measured, respectively, with a barometer (U=0.001 mbar/abs; k=2) and thermometer (U=0.1 o C; k=2). Furthermore, the humidity was measured from a hygrometer (U=3 %HR; k=2). These instruments were calibrated in an accredited metrology laboratory [1]. Tables 1 shows the experimental data. 4 Results and Discussion The analyses of the calibration experimental data were realized according Guide SIM [8]. Different polynomials were assessment in order to selection the best model
Experimental analysis of random errors for calibration 4297 for represent de physical problem [1]. Thus, the results confirmed that polynomial of grade three is the best adjust, i.e.: the best model for the experimental data. Equation 7 show the polynomial: y (x) = 1 10 6 x 3 + 0,0001x 2 + 1,0053x 0,0211 (7) Then, the random error was calculated from the value indicated for the instrument [(x)] and the value adjusted for the polynomial [y (x)]. Then, the objective is to assessment the probability distribution of random error. According to this goal, the figure 2 show the histogram of the experimental data: Figure 2. Histogram associated to experimental data According to figure above, it is possible to infer that the experimental data follow a normal distribution, but, in order to confirmed this hypothesis, a non-parametric test Kolmogorov-Smirnov (K-S) will be applied. In these order of ideas and considering the background theoretically defined in section 2, tables 2 and 3 show the results of parameter associated to K-S test. Table 2. Random error, Z value and area in a normal distribution Random Error Z value (Normal Dist.) Area in a Normal Dist. Lower Limited Upper Limited Frequency Lower Limited Upper Limited Lower Limited Upper Limited -0.375-0.296 18-2.31-1.58 0.0279 0.1154-0.296-0.257 20-1.58-1.21 0.1154 0.1919-0.257-0.217 20-1.21-0.84 0.1919 0.2794-0.217-0.177 20-0.84-0.48 0.2794 0.3558-0.177-0.138 17-0.48-0.11 0.3558 0.3964-0.138-0.098 39-0.11 0.25 0.3964 0.6136-0.098-0.058 56 0.25 0.62 0.3864 0.6705-0.058-0.019 80 0.62 0.98 0.6705 0.7541-0.019 0.021 52 0.98 1.35 0.7541 0.8395 0.021 0.061 60 1.35 1.71 0.8395 0.9083
4298 José Daniel Hernández-Vásquez et al. Table 3. Frequency and Statistics of Test D Absolute and Theoretical Frequency Accumulated Frequency Statistics of Test D fi ei Fi ei D d = D/n 18 33.41 18.00 33.41 15.41 0.04 20 29.25 38.00 62.66 24.66 0.06 20 33.40 58.00 96.06 38.06 0.10 20 29.19 78.00 125.25 47.25 0.12 17 15.52 95.00 140.77 45.77 0.12 39 82.96 134.00 223.73 89.73 0.23 56 108.50 190.00 332.23 142.23 0.37 80 31.96 270.00 364.19 94.19 0.25 52 32.61 322.00 396.80 74.80 0.20 60 26.30 382.00 423.10 41.10 0.11 According to the table above, the statistic of Test D is 0.37 and for the critical value of D (i.e.: Dcrit) for a Dixon Distribution and a confidence level of 95% is 0.410. Then, it possible to conclude that the experimental data follow a normal distribution considering a probability of 95%. Figure 3 confirms that the principal results of this investigation. The blue curve represents the experimental data and the red curve represent the theoretical profile, i.e.: the profile of experimental data for an infinite population. 5 Conclusions Figure 3. Experimental data and normal distribution In the analysis of experimental data obtained from the calibration of a specific measurement instrument, statistical techniques of parameter estimation, confidence intervals and hypothesis testing are applied in a general way. Together, these techniques are part of the so-called parametric statistics. These techniques are based on specifying a distribution form of the random variable and the statistical results derived from the experimental data. A fundamental hypothesis for the development of these models is to assume that the population from which the sample is extracted corresponds to a normal distribution of probability, however, in a large number of
Experimental analysis of random errors for calibration 4299 cases it is not possible to verify this hypothesis, mainly when, starting from the descriptive statistics, the following results are obtained: Sample mean is very distant from the sample median; Coefficient of asymmetry is different from zero; Profile of the histogram is not equal to the profile of a Gaussian curve. In this sense, this research applied the Kolmogorov-Smirnov method to evaluate the probability distribution associated with random errors in a calibration of scales. The consolidated results showed that, despite having a sample of 400 experimental points, it is possible to adjust the random errors to a Gaussian distribution for a confidence level of 95%. Acknowledgements. To the Faculty of Nutrition and Dietetics (Universidad del Atlántico), for to make available Laboratories, Materials and Non-Automatic Weighing Instruments (scales). Thus, it was possible to realize the experiment procedure of the undergraduate project of the Department Mechanical Engineering: Análisis de la confiabilidad metrológica de balanzas analógicas basado en la aplicación de modelos estadísticos no paramétricos. References [1] P. Grimaldo, C. Nuñez, Analysis of the Metrological Reliability of Analogical Balances Based on the Application of Non-Parametric Statistical Models (in Spanish), Department of Mechanical Engineering. Universidad del Atlántico, 2018. [2] B. Nuñez, P. Viloria, E. Diaz, J. González, Fundamentos De Instrumentación Para El Desarrollo De Prototipos Electrónicos; Educosta, 2010. [3] G.R. Pesenti, Rosolem, R.F. Beck, M.G.R. L.A. Soares, G.R. Santos, P.H. Lopes, R.S. Nazari, V.T. Arioli, Remote Monitoring System for Stationary Lead-Acid Batteries, (in Portuguese), SENDI, Rio de Janeiro. 2012. [4] J. Hernandez, G. Pesenti and M. Frota, Metrological evaluation of a stationary remote monitoring battery system, (in Portuguese), III International Congress in Mechanical Metrology, Gramado, RS, Brazil, 2014, 1-6. [5] G. Rangel, Remote Measurement as Strategy to Monitor Stationary Batteries: Case Study in an Eletric Power Substation (in Portuguese), MSc Dissertation, Postgraduate Programme in Metrology (PósMQI), Rio de Janeiro, RJ, Brazil, 2012. [6] S. J. Pappu, N. Bhatt, R. Pasumarthy, A. Rajeswaran, Identifying Topology of Low Voltage (LV) Distribution Networks Based on Smart Meter Data, IEEE Transactions on Smart Grid, 9 (2017), 5113-5122.
4300 José Daniel Hernández-Vásquez et al. https://doi.org/10.1109/tsg.2017.2680542 [7] G. Ferri, V. Stornelli, F. R. Parente, G. Barile, Full range analog Wheatstone bridge based automatic circuit for differential capacitance sensor evaluation, International Journal of Circuit Theory and Applications, 45 (2016), 2149-2156. https://doi.org/10.1002/cta.2298 [8] SIM Inter-American Metrology System, Guidelines on the Calibration of Non- Automatic Weighing Instruments, 2009. Received: September 15, 2018; Published: October 4, 2018