4.1. Background
General considerations for the direct linear regression of
y on
x are presented, where the
y-dependent coordinate represents the signal in mV/V and the
x-independent variable represents the torque in kN·m. The linear or nonlinear functions are described by the polynomial expansion of a Taylor series
, where the measurements of the variable
are carried out with the associated uncertainties
and
, while the coordinate
is determined with an uncertainty close to zero. These considerations lead to the use of the weighted least squares (WLS) method, which minimises the criterion function (
9) [
37]:
where the parameters of the function are defined by an
dimensional vector
. It should be noted that Equation (
9) does not take into account the correlations between the coordinates, as it is assumed that there are no correlations between coordinates during the measurements.
The condition for minimising Equation (
9) implies that the gradient of the criterion function is zero, meaning that the first derivatives with respect to all parameters are equal to zero, i.e.,
0. This leads to obtaining
linear equations with respect to the parameters of the vector
:
and
.
The above linear equations, which depend on the parameters
, can be solved using the determinant method. This is because they can be expressed in the following form:
where
4.3. Results of Fitting to a Straight Line and 3-Degree Polynomial
As a result of the torque calibration measurements, the transducer signals listed in
Table 1 and
Table 2 were obtained [
27].
On the basis of
Section 4.2.1, straight line fitting is carried out. The obtained characteristics are shown in
Figure 6, where measurement points are marked in the form of squares, points are measured in the range from −1100 kN·m to 1100 kN·m, and two fitted linear characteristics above the calibration range are plotted using the weighted method of least squares (WLS) (dashed blue line for anticlockwise torque and dashed red line for clockwise torque). The equations of both matched straight lines are also given.
In order to show the surroundings of the measuring point and the line adjustments from
Figure 6, the last calibration measurement point for clockwise torque is included in the magnified area in
Figure 7 below. For the selected calibration point, i.e., torque = 1100 kN·m, the area from
Figure 6 has been enlarged, where the WLS line is adjusted, marked with a red dash-dot line. You can also see two pairs of curved lines determining a narrower coverage region in the case of using the MCM method (red dotted lines and a wider one in the form of blue dashed lines for estimating the corridor using the law of propagation of uncertainty). The following surrounding corridor coverage for the selected measured point looks analogous for all other points, including points measured in the anticlockwise range.
Figure 6.
Straight-line adjustment of the relationship of the tared transducer signal as a function of torque from the WLS method. Data from
Table 1,
Table 2 and
Table 3.
Figure 6.
Straight-line adjustment of the relationship of the tared transducer signal as a function of torque from the WLS method. Data from
Table 1,
Table 2 and
Table 3.
In addition, an extrapolation of the linear characteristics of the tared transducer signal was performed as a function of torque in the range from −100 MN·m up to 100 MN·m. The results are illustrated in
Figure 8.
Figure 8 includes linear extrapolations of straight line fitting in the range from −1100 kN·m to 1100 kN·m of the torque to the range from −100 MN·m to 100 MN·m of the torque in both the anticlockwise and clockwise directions. In addition, there is a table containing selected characteristic points with extended uncertainties of the signal in steps of 0.5 MN·m in the range from −5 MN·m to 5 MN·m of the torque.
In order to check the nonlinearity, the adjustment to the measured coordinates was achieved with cubic splines, i.e., polynomials of the third degree, which, due to their characteristics, can adapt to various types of nonlinearity. Then, a cubic spline curve fit is performed. The expanded uncertainties of the signal
are obtained by transforming it from Equations (
21) and (
22). The results are shown in
Table 4,
Table 5 and
Table 6.
Figure 9 and
Table 4,
Table 5 and
Table 6 shows the full fitting characteristics of the cubic spline.
Figure 9, on the other hand, shows two matching cubic characteristics in the calibrated range with measurement points, allowing for possible nonlinearities in the characteristics of the tested transducer. In terms of calibration, formulas describing these curves with cubic splines in both measured directions are given. The penultimate calibration point for the clockwise direction is also indicated, whose enlarged surroundings are shown in
Figure 10. It is easy to see that the coverage corridor is broader for both the MCM method and the method of law of propagation of uncertainty (LPU) compared to the coverage corridor for the linear fit (
Figure 7). As before, for linear fitting, the MCM method determines a narrower corridor than the LPU method. The following surrounding corridor coverage for the selected measured point looks analogous for all other points, including points measured in the anticlockwise range.
In addition, a cubic spline curve was fitted to the three parameters for the data in
Table 1, and the resulting equation for the curve is
.
For anticlockwise signal measurements, a cubic spline curve was fitted to the three parameters for the data in
Table 2, and the resulting equation for the curve is
.
In
Figure 11, the absolute expanded uncertainties, and in
Figure 12, the relative expanded uncertainties determined by the LPU method and MCM are presented.
Figure 11 shows the characteristics of half of the corridor of the expanded absolute uncertainty of the signal for linear and cubic spline fitting in both directions for both the MCM and LPU methods. These characteristics indicate a nonlinear increase in the expanded uncertainty of the signal with an increase in the measured torque setpoint. Due to the anticlockwise and clockwise measurement directions, the characteristics of the measured transducer are not symmetrical, and there is a nonzero uncertainty in the signal for a zero torque. The smallest extended uncertainties occur for MCM estimation for linear fitting, slightly higher for LPU estimation, and for spline matching in the measured range, a significant increase in signal uncertainty is observed, in particular for the LPU method.
Table 4.
Adjustment for clockwise signal measurements, adjustment factors , , , . The expanded absolute uncertainties and relative uncertainties for 0.95 probability of signal—coverage factor for 0.95 probability of signal .
Table 4.
Adjustment for clockwise signal measurements, adjustment factors , , , . The expanded absolute uncertainties and relative uncertainties for 0.95 probability of signal—coverage factor for 0.95 probability of signal .
Torque [kN·m] | Expanded Uncertainty of Signal Monte Carlo Method MCM | Expanded Uncertainty of Signal—The Law of Propagation of Uncertainty LPU |
---|
| Absolute
[mV/V] | Relative
[%] | Absolute
[mV/V] | Relative
[%] |
---|
100 | 0.000041 | 0.159 | 0.000058 | 0.225 |
200 | 0.000040 | 0.077 | 0.000056 | 0.109 |
300 | 0.000050 | 0.064 | 0.000071 | 0.091 |
400 | 0.000053 | 0.051 | 0.000076 | 0.073 |
600 | 0.000084 | 0.054 | 0.000119 | 0.076 |
800 | 0.000113 | 0.055 | 0.000161 | 0.077 |
1000 | 0.000121 | 0.047 | 0.000171 | 0.066 |
1100 | 0.000193 | 0.068 | 0.000274 | 0.096 |
Table 5.
Adjustment for anticlockwise signal measurements, adjustment factors , , , and , . Expanded absolute uncertainties () and relative uncertainties —coverage factor for probability of the signal .
Table 5.
Adjustment for anticlockwise signal measurements, adjustment factors , , , and , . Expanded absolute uncertainties () and relative uncertainties —coverage factor for probability of the signal .
Torque [kN·m] | Expanded Uncertainty of Signal Monte Carlo Method MCM | Expanded Uncertainty the Signal—The LAW of Propagation of Uncertainty LPU |
---|
| Absolute
[mV/V] | Relative
[%] | Absolute
[mV/V] | Relative
[%] |
---|
−1100 | 0.000188 | 0.066 | 0.000267 | 0.093 |
−1000 | 0.000188 | 0.046 | 0.000169 | 0.065 |
−800 | 0.000106 | 0.051 | 0.000151 | 0.073 |
−600 | 0.000080 | 0.051 | 0.000134 | 0.073 |
−400 | 0.000045 | 0.043 | 0.000063 | 0.061 |
−300 | 0.000038 | 0.049 | 0.000054 | 0.070 |
−200 | 0.000029 | 0.056 | 0.000041 | 0.079 |
−100 | 0.000030 | 0.114 | 0.000042 | 0.161 |
Table 6.
Table of parameters of 3-degree curves —uncertainties and correlations coefficients obtained by the MCM.
Table 6.
Table of parameters of 3-degree curves —uncertainties and correlations coefficients obtained by the MCM.
| Anticlockwise | Clockwise |
---|
| | |
| | |
| 0.00025956 | 0.00025968 |
| | |
| | |
| | |
| | |
| | |
| −0.987 | −0.988 |
| 0.918 | 0.926 |
| −0.796 | −0.791 |
| −0.967 | −0.971 |
| 0.864 | 0.856 |
| −0.947 | −0.937 |
Figure 11.
Plot of absolute expanded uncertainties using LPU and MCM for the straight line fit and cubic spline fit—signal values near zero are not included in the graph because the concept of relative error loses its meaning. Data from
Table 3,
Table 4 and
Table 5.
Figure 11.
Plot of absolute expanded uncertainties using LPU and MCM for the straight line fit and cubic spline fit—signal values near zero are not included in the graph because the concept of relative error loses its meaning. Data from
Table 3,
Table 4 and
Table 5.
Figure 12.
Plot of relative expanded uncertainties using LPU and MCM for the straight line fit and cubic spline fit—signal values near zero are not included in the graph because the concept of relative error loses its meaning. Data from
Table 4,
Table 5 and
Table 7.
Figure 12.
Plot of relative expanded uncertainties using LPU and MCM for the straight line fit and cubic spline fit—signal values near zero are not included in the graph because the concept of relative error loses its meaning. Data from
Table 4,
Table 5 and
Table 7.
Figure 12 contains expanded relative uncertainties, which, for both matches and both methods below −300 kN·m and above 300 kN·m of the calibrated range up to −100 kN·m and above 100 kN·m, are from 0.03% to a maximum of 0.08%. Near torque values close to zero, the relative uncertainty of measuring a virtually close-to-zero signal loses its meaning. It also turned out that extrapolated expanded uncertainties when fitting using the cubic spline method increase significantly relative to the extrapolated expanded uncertainties for a linear fit.
Table 8 presents the relative expanded measurement uncertainty for the 5 MN·m torque transducer designed to measure clockwise and anticlockwise torque calculated by different methods. Finally, taking into account only the linear nature of the transducer,
Table 8 gives the estimated values of expanded relative uncertainty in both the calibrated and extrapolated ranges, to an absolute torque value of |MK| = 5 MN·m. These values are below 0.06% for the LPU method and 0.05% for the MCM method in both directions of measurement.