0 evaluări0% au considerat acest document util (0 voturi)
65 vizualizări9 pagini
This paper proposes a non-traditional paradigm that considers measurement to be a process. Statistical process control allows a manufacturer to confirm that a process is operating "in control" the results are used to provide estimated uncertainty values associated with the measurement result.
This paper proposes a non-traditional paradigm that considers measurement to be a process. Statistical process control allows a manufacturer to confirm that a process is operating "in control" the results are used to provide estimated uncertainty values associated with the measurement result.
This paper proposes a non-traditional paradigm that considers measurement to be a process. Statistical process control allows a manufacturer to confirm that a process is operating "in control" the results are used to provide estimated uncertainty values associated with the measurement result.
Comit Interorganismos de Ductos de Petrleos Mexicanos
Congreso Internacional de Ductos
Mrida, Yucatn, 14 16 Noviembre del 2001 Paper ID: OAM-18 STATISTICAL CONTROL OF THE MEASUREMENT PROCESS
Thomas Kegel Colorado Engineering Experiment Station, Inc. Nunn, Colorado, USA www.ceesi.com (970)897-2711
ABSTRACT This paper proposes a non-traditional paradigm that considers measurement to be a process. In a manufacturing process, it is recognized that many small effects can contribute to random variance in the process output. These effects arise from an operator, a procedure, the environment, and raw materials as well as the machinery itself. The economic equilibrium between the cost of identifying the source of an effect and the benefit realized from reducing the impact inevitably results in some level of random variance. The manufacturing community has developed a statistical tool, statistical process control (SPC), to monitor the consistency of the random variance. The application of SPC allows a manufacturer to confirm that a process is operating in control and therefore producing consistent products. From a statistical point of view, the measurement and manufacturing processes are very similar. The result of a measurement will exhibit random variance due to the effects of operators, procedures, and the environment as well as the instrument itself. The SPC algorithms are therefore equally applicable to the measurement process. This paper describes by example the application of statistical process control to several calibration processes. In addition to monitoring consistency, the results are used to provide estimated uncertainty values associated with the measurement result. INTRODUCTION As technology continues to play an increasing role in todays energy business, the demand for supporting measurements also increases. In addition, competitive pressures to continuously improve processes further expand the role of measurement. These economic forces highlight the importance of understanding measurement science and the implications of decisions made based on measurement results. This understanding is even more critical given the increasing value of the commodity being measured. In the manufacturing industries the correlations between measurement, statistics and process improvement have been well established for many years 1,2 . The automotive industry in particular 3 has realized benefits in productivity and quality as a result of improved measurement science. More recently there has been a realization that the same statistical techniques that have been developed for manufacturing are applicable to measurement science 4,5 . The underlying statistics are incapable of differentiating between the geometry of a manufactured part and the result of a measurement process One of the more important tools developed by the manufacturing community is Statistical Process Control. It is usually presented in the form of a control chart designed to interpret the underlying statistics in a simplified form. The simplification allows for reliable decision making by practitioners without extensive training in statistics. This paper provides an introduction to the application of statistical analysis to measurement science. THE MEASUREMENT PROCESS When a measurement is made there is a tendency to consider only the performance of the measuring instrument itself. When stating the uncertainty associated with a measurement, the instrument manufacturers specifications are often given as the only source of uncertainty. In fact, the measurement is a process that includes the instrument as only one of several components. Other components that potentially contribute uncertainty to the process output include: an operator, operating procedures, instrument maintenance and the environment. Each of these components contribute potentially small random effects that when combined result in a larger random effect associated with the measurement process output. The sources of the small random effects are generally unknown and potentially costly to identify. In fact, as long as the process is capable of making the required measurement, there is no need to identify the sources of random variation. It is only Comit Interorganismos de Ductos de Petrleos Mexicanos required that the process be monitored to ensure that the random effects are consistent over time. For the practitioner, the brief discussion above is summarized in two important questions to be asked regarding a measurement process. First, is the uncertainty appropriate for the application? Second, are measurements consistent over time? It has been proposed 4 that if the answer to either of these questions is no then the process is not capable of making a measurement. This paper presents statistical tools to allow the practitioner to answer these questions and ensure valid measurement. A simple example has been formulated to demonstrate the statistical principles. The subject is a fictitious pressure transducer that is used over a 1001000 psi range. The output is a voltage that varies over the 05 volt range, the nominal sensitivity is therefore 200 psi/volt. The transducer is calibrated every 90 days, a history has been developed based on 16 calibrations made between January 1999 and December 2002. Each calibration involves determining the transducer sensitivity based on values calculated from pressure and voltage standards. Sensitivity data are obtained at ten pressure levels equally spaced over the input range. The calibration results from December 1999 are shown in Figure 1. These data would appear to suggest a change in sensitivity over the range of the transducer, the solid line represents a proposed correlation. The magnitude of the change in sensitivity is 1.9 [psi/v] from 100 psi to 1000 psi. Is this significant? Is it typical of the calibration history, or a characteristic of this particular calibration? Some decisions need to be made based on these results. Does the change of sensitivity affect measurements made with the transducer? Should the transducer be adjusted? Figure 2 shows the December 1999 data (closed circles connected by lines) compared with the entire calibration history (open circles). The data at each pressure level are distributed over a range of sensitivity values, this is indicative of random effects. It appears as if the 1.9 [psi/v] change in sensitivity lies within the limits of these random effects. Are the data in Figure 1 the result of a series of random effects rather than a trend? STATISTICAL PROCESS CONTROL The scenario described above does not provide the practitioner with adequate basis on which to answer some potentially important questions. The calibration data are adequate but the proper tools are not available. The reliability of the required decisions can be improved through the use of Statistical Process Control (SPC). The heart of SPC is the control chart. It assists the practitioner to quickly evaluate to status of a calibration history and confidently make the correct decision. What is a control chart and how is it developed? This process is now described: The ten sensitivity values that make up a calibration are used to calculate a mean and standard deviation. These values, when plotted against time make up the two types of control charts. The x chart contains the mean values while the standard deviations are plotted in the s chart. Control charts for the current example are shown in Figures 3 and 4. The symbols represent the mean and standard deviation values of sensitivity calculated for the 16 calibrations. The solid lines represent control limits, they are calculated from values of accumulated standard deviation 4 . The measurement process is considered to be in a state of statistical control if a data point lies between the control limits of the x chart and below the single control limit of the s chart. Only one limit is included in the s chart because the values for standard deviation are always greater than zero. The control limits in the charts are not fixed because they are updated as calibration data are accumulated. The term statistical control means that the measurement process is operating in a consistent manner. The control limits quantify the level of consistency. If a process is operating consistently, and nothing is done to change that process, then there is a level of confidence that the consistent performance will continue into the future. An out of control condition observed in a control chart is an indication of a problem with the measurement process that needs to be investigated. The great value in the use of control charts is in separating potentially serious measurement problems from naturally occurring random variations. In the present example, the control charts indicate that the apparent trend in Figure 1 is nothing more than random effects and does not warrant adjusting the transducer. The term level of confidence appears in the discussion above. It accounts for the fact that statistical analyses never produce absolutely conclusive results. Conclusions are always stated with some level of confidence. The particular level depends on the application, standard practice in measurement science it is a 95% level of confidence. The shape of the control limits of Figure 3 is a graphical example of the confidence interval concept. The x chart shows a steady decrease in the control limit interval width, decreasing from 2.46 [psi/v] in April 1999 0.67 [psi/v] in December 2002. As data are accumulated the level of confidence associated with an in control conclusion increases. This increase in confidence results in the control limits moving closer together. The s chart would show a similar trend except that the standard deviation value for the first data point is small enough (relative to subsequent data) to offset the low initial level of confidence. In an industrial SPC application the control charts are maintained based on samples of the manufacturing process. In a measurement process application periodic samples of the measurement are made, these samples are called calibrations. While the control charts are maintained based on calibration data (samples), the relevant application is what takes place between the calibrations. The validity of the industrial control charts is dependent on how well the sample represents an entire batch of the product. Similarly, the validity of the measurement process control charts depends on the degree of measurement and calibration process similarity. ESTIMATING UNCERTAINTY Two values of standard deviation, with different interpretation, are calculated from the calibration data to form the control limits. A pooled value of calibration standard Comit Interorganismos de Ductos de Petrleos Mexicanos deviations, called s w for within, defines the control limit for the s chart (Figure 4). The term pooled refers to a process for combining multiple standard deviation values. The pooled value represents the process variation observed in the time required to perform a ten point calibration. The second standard deviation, calculated from the mean values, defines the control limits of the x chart (Figure 3). This standard deviation, called s b for between, represents random effects that are only observed over long periods of time. The duration of a long period of time depends on the application, in the present example the duration is four years. The reported standard deviation, s r , accounts for both short term and long term random effects. The measurement uncertainty will be underestimated both effects are not included. The value for s r is calculated from values of s w and s b
combined in quadrature (root sum square): 08 . 1 03 . 1 32 . 0 s s s 2 2 2 w 2 b r = + = + = [psi/v] The interpretation of s r is as follows: all of the random effects associated with the measurement process amount to (21.08) = 2.16 [psi/v] with a confidence level of 95%. The 2 term means that two standard deviation values are required to achieve the 95% level of confidence. The s r value can be confirmed by inspection of Figure 2; 95% of those data points fall within an interval with width equal to 2.16 [psi/v]. While invaluable in the determination of uncertainty, the control charts cannot identify the systematic effects that contribute uncertainty. What is a systematic effect? A random effect will contribute to observed changes in a repeated measurement while a systematic effect will not change through repetition of the same measurement. A typical systematic effect present in the current example would be the uncertainty associated with the standard used to calibrate the transducer. When using control chart data in an uncertainty analysis, uncertainty estimates attributed to relevant systematic effects must be determined by other methods. PREDICTING THE FUTURE When an instrument is calibrated, the purpose of the calibration is to predict the future. While the present calibration of an instrument can be used to re-evaluate events in the past, the most common application of the result is in future events. The most recent calibration results for the fictitious pressure transducer are used for the proceeding 90 days. While predicting the future with absolute certainty cannot be accomplished, statistical tools can be applied to estimate the uncertainty in the prediction. The mean values for the first four calibrations are contained in Figure 5. A simple linear fit is made of mean values and calibration date (the date is expressed as the number of days since January 1, 1900). The solid lines represent the prediction interval, the interval within which a future calibration can be predicted to fall 6 . The interval width increases over time, from 1.7 [psi/v] in September 1999 and 8.4 [psi/v] in December 2002. The shape of the prediction interval is the result of the range of slope and intercept values describing lines that could be fitted to the four calibration points. Stated a different way: when a line is fitted to data there in uncertainty associated with both slope and intercept, this results in the increase in interval width when the data are extrapolated. The prediction interval is fundamentally different from the control limits despite similar appearances. While new control limit values are calculated for each calibration the values for past calibrations remain the same. The prediction interval is recalculated for the entire calibration history when a new data point is obtained. The important conclusion to be drawn from Figure 5 is that uncertainty grows over time after a calibration. Figure 6 shows the prediction intervals resulting from six (dashed line) and ten (solid line) data points. The interval width decreases as additional data points are obtained. The December 2002 values are reduced from 8.4 [psi/v] to 1.2 [psi/v] as the data points increase in number from four to ten. While useful as a learning aid, predicting instrument performance several years into the future is not a realistic application. A more practical need would be to predict the uncertainty growth from one calibration to the next. This information is contained in Figure 7. The abscissa is the number of accumulated calibrations. The ordinate is the predicted increase in interval width from the last calibration to the next. For example, if a history consists of eight calibrations, the prediction interval width is expected to increase by 6.5% by the time of the ninth calibration. The linear fit determined from the first four data points (Figure 5) predicts a change in sensitivity of approximately 1.0 [psi/v] per year. On the one hand, this behavior should prompt an investigation of the transducer because it might indicate measurement problem. On the other hand, the trend could be nothing more than random values appearing in a particular order. The data of Figure 1 appeared to have indicated a trend that wasnt present, the same thing may be happening with these data. The results in Figure 6 indicate that the trend over time is not predicted as more data are accumulated. The results are summarized in the plot of Figure 7, as additional calibration data are obtained the slope approaches a value of zero. SPC APPLICATIONS The concepts of SPC have been applied to a number of applications in a flow calibration facility. The earliest application involves pressure instrumentation that is maintained in a regular calibration program 7,8 . Analyses have been performed on both laboratory and industrial grade transducers. The control charts are used to ensure measurement consistency and estimate uncertainty. A more recent application involves the use of flowmeter check standards 9,10,11 . Several flowmeters have been permanently installed in series with the test section of a calibration facility. One of these meters is always calibrated at the same time as the meters under test (MUT). This application differs from the first in that the flowmeters are not used as standards. Control charts from one of the meters are contained in Figures 8 and 9. The abscissa values differ from those of Figures 3 and 4. The flowmeter performance varies with the Comit Interorganismos de Ductos de Petrleos Mexicanos input, the random effects increase as the flowrate decreases. In order to compare performance at different flowrates, the data are normalized. The control charts in Figures 8 and 9 are based on 3976 data points contained in 224 calibrations obtained during a six month period of time. Calibration data obtained before and after this time interval were based on different check meters and are summarized in different control charts. The check standard concept is very valuable because a meter with a well established history is always subject to the same conditions as a MUT. When the check meter indicates that the calibration process is in control this means that the current calibration of the MUT is consistent with the calibration history of the check standard. The control chart data also form an important part of the uncertainty estimate for the facility, the value of s r is assumed to account for all random effects present during the MUT calibration. The ability to apply the estimated uncertainty to the MUT is enhanced by the calibration history of the check standard. The occasional data points that exceed the control limits in Figures 8 and 9 are not a problem because the limits are defined based on a 95% level of confidence. At this level of confidence 5 points out of 100 can fall outside the limits and the process will remain in control. In general, a decision based on one data point in the middle of a group is relatively straightforward, hindsight is always 20/20. For example, the control charts clearly show that the data from mid September to mid October are in control. A practitioner looking at the current control charts in early September does not have this information. They are faced with an out of control condition (OCC), what does this mean? Is the current calibration one of the 5% that can be OCC?, Is the process truly out of control? There are a number of recommended steps to react to an OCC: The first step is to confirm OCC indication from both control charts, most of the OCC cases in Figures 8 and 9 appear only on one chart. The second step is to wait for a few calibrations and see if the OCC indicator persists. Referring again to Figures 8 and 9, most of the OCC indicators exist only for one or two data points. Third, the manufacturing community has developed a variety of rules to interpret control chart trends and patterns 1 . Finally, an engineering investigation of the process should be undertaken with the intent of localizing the source of the OCC. CONCLUSION The application of traditional statistical analysis tools to measurement science has been briefly described. This was accomplished based on a simple example as well as some real world data. Through the use of Statistical Process Control, the practitioner has a tool to help answer the following questions: 1) Are the measurement process results consistent? 2) What is the uncertainty due to random effects? 3) How much uncertainty growth will occur prior to the next calibration? REFERENCES 1. Juran, J. M., Quality Control Handbook, McGraw-Hill, 1951. 2. Bothe, David R., Measuring Process Capability, McGraw- Hill, 1997. 3. Automotive Industry Action Group, Measuring System Analysis Reference Manual, Southfield, MI, 1990. 4. Croarkin, Carroll, Measurement Assurance Programs, Part 2: Development and Implementation, NBS Special Publication 676-2, 1985. 5. Castrup, H. T., et al, Metrology-Calibration and Measurement Process Guidelines, NASA Reference Publication 1342, 1994. 6. Hahn, G. J. and Meeker, W. Q., Statistical Intervals, A Guide for Practitioners, John Wiley, 1991. 7. Kegel, T. M., "Statistical Control of a Pressure Measurement Process," Transactions of the ISA: Journal of the Instrument Society of America, 1996. Vol 35, p. 69-77. 8. Kegel, Thomas, Statistical Control of a Differential Pressure Instrument Calibration Process, 45 th
International Instrumentation Symposium, Albuquerque, New Mexico, May 2-6, 1999. 9. Kegel, Thomas, Uncertainty Issues Associated With a Very Large Capacity Flow Calibration Facility, Measurement Science Conference, Anaheim, California, January 2021, 2000 10. Kegel, T. M., Statistical Control of a Flowmeter Calibration Process, 46 th International Instrumentation Symposium, Bellevue, WA, April 30 May 5, 2000. 11. Kegel, T. M., Long Term Ultrasonic Meter Performance Data, AGA Operations Conference, May, 2001. Comit Interorganismos de Ductos de Petrleos Mexicanos
197 199 201 203 0 200 400 600 800 1000 Applied Pressure [psia] S e n s i t i v i t y
[ p s i / v o l t ] 197 199 201 203 0 200 400 600 800 1000 Applied Pressure [psia] S e n s i t i v i t y
[ p s i / v o l t ] Figure 1: December 1999 Calibration Data Figure 2: Entire Calibration History (December 1999 data shown in black) Comit Interorganismos de Ductos de Petrleos Mexicanos
197 199 201 203 Nov-98 Nov-99 Nov-00 Nov-01 Nov-02 Date M e a n
[ p s i / v ] 0.4 0.8 1.2 1.6 2.0 Nov-98 Nov-99 Nov-00 Nov-01 Nov-02 Date S t a n d a r d
D e v i a t i o n
[ p s i / v ] Figure 3: Control Chart of Mean Data ( x chart) Figure 4: Control Chart of Standard Deviation Data (s chart) Comit Interorganismos de Ductos de Petrleos Mexicanos
184 188 192 196 200 204 208 Nov-98 Nov-99 Nov-00 Nov-01 Nov-02 Date M e a n
[ p s i / v ] 196 198 200 202 204 Nov-98 Nov-99 Nov-00 Nov-01 Nov-02 Calibration Date M e a n
[ p s i / v ] Interval width based on: 6 points 10 points Figure 6: Prediction Intervals Based on Six and Ten Calibrations Figure 5: Prediction Interval Based on Four Calibrations Comit Interorganismos de Ductos de Petrleos Mexicanos
0 5 10 15 20 25 4 6 8 10 12 14 16 Calibrations U n c e r t a i n t y
G r o w t h
[ % ] -1.2 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 L i n e a r
D r i f t
[ p s i / v / y r ] Growth Drift 0.0 0.4 0.8 1.2 1.6 Jun-00 Aug-00 Oct-00 Dec-00 Date S t a n d a r d
D e v i a t i o n
[ s ] Figure 7: Uncertainty Growth and Linear Drift Figure 8: Control Chart for Standard Deviation Data from a Flowmeter Check Standard Comit Interorganismos de Ductos de Petrleos Mexicanos
-2.0 -1.0 0.0 1.0 2.0 Jun-00 Aug-00 Oct-00 Dec-00 Date M e a n
[ s ] Figure 9: Control Chart for Mean Data from a Flowmeter Check Standard
Mental Math: How to Develop a Mind for Numbers, Rapid Calculations and Creative Math Tricks (Including Special Speed Math for SAT, GMAT and GRE Students)