Detection of Non-Random Patterns in Shewhart Control Charts: Methods and Applications A. Rakitzis and S. Bersimis Abstract- The main purpose of this article is the development and the study of runs rules applied to a Shewhart type control chart for the mean in order to detect non-random patterns in Phase I, when we are interested to conclude if the process is in a stable condition for the first twenty five or thirty samples in order to estimate the process parameters. A characterization scheme such as which rule is better for which pattern, is given. This characterization is extremely useful to the practitioners. A few applications of the proposed methodology are presented. Index Terms- Quality Control; Statistical Process Control; Runs Rules, X-bar Chart, non-random patterns. I. INTRODUCTION Statistical process control techniques are widely used in industry. The most common process control technique is control charting. Statistical based process control charts are tools for controlling and monitoring a manufacturing process. Their goals are to detect shifts (Shewhart type control charts) or / and drifts (CUSUM and EWMA type control charts) in the mean level or the dispersion of the process. These shifts or / drifts in the mean level or the dispersion of the process are giving us evidence that the process is out-of-control. The term shift corresponds to a sudden and of high level change in the mean or the variance of a process quality characteristic. In contrast, the term drift corresponds to a gradual and small change in the mean or the variance of a process quality characteristic. Additionally, a possible out-of-control condition in the monitored process, steams from the presence of nonrandom patterns in the sequence of the values of the monitored variable, like Cycles, Trends, Small Shifts, Stratification and Mixtures. There are two distinct phases of control charting, Phase I and Phase II. In Phase I, charts are used for retrospectively testing whether the process was in control when the first subgroups were being drawn. In this phase, the charts are used as aids to the practitioner, in bringing a process into a state of statistical in-control. Once this is accomplished, the control chart is used to define what is meant by statistical in-control. In Phase II, control charts are used for testing University of Piraeus Department of Statistics and Insurance Science, Greece whether the process remains in control when future subgroups are drawn. In Phase I, Shewhart control charts are very effective, since they are easy to construct and the interpretation is clear. Specifically, the meaning of a pattern appearing in a Shewhart type control chart is straightforward and has a physical meaning. A Shewhart type control chart is a graphical display of a process or a product quality characteristic that has been measured or computed from a sample versus the sample number or time. The basic characteristics of a univariate Shewhart process control chart are the «Center Line» (C.L), the «Upper Control Limit» (U.C.L), and the «Lower Control Limit» (L.C.L). The ordinary rule applied to a Shewhart type control chart for declaring a possible out-of-control condition in a manufacturing process is the occurrence of a point outside the control limits. For example, in Figure, an out-ofcontrol process is presented. As we may easily observe an out-of-control signal is given at the time 95 th sample. 3 3 5 5 7 7 9 9 Measurement C.L U.C.L. L.C.L Figure : A classic Shewhart Type X control chart In addition, many sensitizing run rules have been suggested for the Shewhart control charts in order to make them more sensitive in the detection of drifts in the mean of the process. From the above mentioned run rules, the need for additional limits, named warning limits, arises. Thus, we supplement the Shewhart type control chart with an «Upper Warning Limit» (U.W.L), and a «Lower Warning Limit» (L.W.L). These runs rules were first proposed by the Western Electric Company [9]. Additional run rules have been suggested by Page [], Roberts [7], Bissel [], Nelson [5], Klein [], Khoo [3], and Rakitzis [].
For example, a classic sensitizing rule based on runs is the well known fourth rule of the runs rules given by the Western Electric Company [9]. The rule is eight consecutive points plot on one side of the center line. In Figure the same process as in Figure is given. Using the runs rule just mentioned we may easily see that the out-ofcontrol condition is detected at the time th sample. 3 3 5 5 7 7 9 9 Measurement C.L U.W.L.(s) L.W.L.(s) U.C.L. L.C.L Figure : A Shewhart Type X control chart with runs rules The use of sensitizing rules in a control chart results in the increase of the sensitivity of the chart for detecting shifts and/or trends. The most famous measure of the performance of a chart (in favor of sensitivity) is the Average Run Length (ARL). ARL is defined as the expected number of points that are plotted in the chart until we get an out-of-control signal. In order to compare different control charts, they must have the same in-control ARL, the value of ARL for an in-control manufacturing process, and compare their out-of-control ARL values for a specified shift in the mean of the process. But, it is obvious that ARL as a measure of performance is useful only in Phase II, where we monitor on-line the manufacturing process. In Phase I, we use the retrospective data for the estimation and calculation of the control limits. Although the use of sensitizing rules is usual in Phase II, Montgomery [] suggests the use of them, not in Phase II as a solution to the problem of non-sensitivity of a Shewhart chart but instead in Phase I, in order to recognize nonrandom patterns in the retrospective data, characterize them and take the necessary actions so as to remove them. These actions will result in a better estimation of the control limits As already was declared the main purpose of this article is the development and the study of run rules applied to a Shewhart type control chart for the mean in order to detect non-random patterns in Phase I. Specifically, in Section, we discuss some methodological aspects of our study and some potential for further theoretical derivations. Furthermore, in Section 3, we present analytically the findings of an extensive numerical study. Finally, in Section we point out some concluding remarks and topics for further research. II. METHODOLOGY In this Section we present the basic ideas of our methodology. As we already stated our primary goal is to study the well known Shewhart type X control chart in Phase I, when this chart is supplemented with sensitizing run rules. The basic idea is to supplement this very chart with appropriate runs rules in order to discriminate an in-control process from an out-of-control process. Furthermore, we propose additional rules for the Phase I study of a process using an X control chart in order to identify an out-of-control condition which stems from the presence of non-random patterns in the sequence of the values of the monitored variable, like Cycles, Trends, Small Shifts, Stratification and Mixtures. Examples of nonrandom patterns are given in Figures 3,, and 5. 3 3 5 5 7 7 9 9 Measurement C.L U.C.L. L.C.L Figure 3: A trend in the process 3 3 5 5 7 7 9 9 Measurement C.L U.C.L. L.C.L Figure : Cyclical Pattern in the process 3 3 5 5 7 7 9 9 Measurement C.L U.C.L. L.C.L. U.W.L. L.W.L. U.I.W.L. L.I.W.L. Figure 5: Stratification Pattern In Table I, we present the main characteristic of each non random pattern, while in Table II we give corresponding image of each non random pattern on the control chart. Specifically, in this article, we propose a set of appropriate rules for each case. These rules are not based in the first appearance of a specific pattern (this is usually the
case in Phase II analysis). The new rules that are proposed in this article are based on the number of runs (in general on the number of scans and specific patterns) that are present in the sequence. Non-Random Pattern Characteristic Shifts A sudden huge shift. Trends or Drifts A continuous small drift. 3 Cycles A continuous repeated movement. Stratification Almost constant values. 5 Mixtures No points fluctuate near the C.L. Table I: Characterization of Non-Random Patterns Non-Random Pattern Control Chart Shifts Points near and/or beyond the control limits Trends or Series of points going up or down Drifts from the C.L.(in one direction) 3 Cycles Patterns of points representing cycles Stratification Points concentrate in zone near by C.L. 5 Mixtures Big distances between consecutive points or consecutive points lie in opposite positions near in C.L. Table II: Image of Non-Random Patterns on the chart The use of runs appears to be a good solution in the problem of detecting non-random patterns, taking into account that runs are the usually used statistic in randomness test. The X control chart may be seen as a sequence of values of a statistic. Thus, in Phase I we may apply runs rules testing a hypothesis similar in nature with the well known randomness tests, as well as an additional hypothesis that the sequence of values come from a common distribution. We are interested to determine if the process is in a stable condition for the first twenty five or thirty samples in order to estimate the process parameters. A characterization scheme such as which rule is better for which pattern, is given. This characterization is extremely useful to the practitioners. A few applications of the proposed methodology are presented. III. NUMERICAL STUDY In this section, we present the main findings of an extended numerical study. Furthermore, in the beginning we give tables (Tables III, IV, and V) with critical values for the statistics used, that are M, and B n,w,k. These values obtained for a pre-specified probability p, which corresponds to the probability that the statistic X be plotted inside one appropriately selected zone under the assumption that the process is under a stable state, as well as for a sequence of a pre-specified length equal to m (the number of points plotted on the control chart). Both these m k statistics were used for the identification of the nonrandom patterns such as cycles, trend and stratification. k c Exact.9 3. 5.5 5 3.5 Table III: Detecting Cycles using M, for m = 3, n = 5 n k k c Exact 9.5 3 5.537.53 5.5 Table IV: Detecting Stratification using m = 3, n = 5 M n, k for w c Exact 5 7.377593 7.39573.3979.773 Table V: Detecting Trend using B n,w,k for m = 3, n = 5 The numerical study has been planned as following: A: An in-control process was generated and the proposed rules were applied on the process in order to record the times that a run rule faulty detected a non-random pattern. The number of the false alarms was recorded, and B: An out-of-control process was generated having nonrandom patterns according to a specific model and the proposed rules were applied on the process in order to record the times that a run rule succeeded to detect the nonrandom pattern. The number of the right detections was recorded. The above procedure was repeated times. The models that were used in order to generate processes having non-random patters are the following: C: Cyclical patterns require the presence of a periodic movement and a specific structure, of length equal to the period, for this movement in terms of the variance of the statistic. C., with cycle period equal to and structure (,,,3,,,,-,-,-3,-,-), C., with cycle period equal to and structure (,.5,,,,.5,,-.5,-,-,-,-.5), C.3, with cycle period equal to and structure (,,.5,,,-,-.5,-). C-Rule: For detecting one cyclical pattern in the sequence of the values of the statistic on the control chart we use one rule, which is based on the number of overlapping runs of length k in the one side of the C.L. S: Stratification patterns require random or non random movement of the points too close (with too small variance) to the center line of the control chart. S., with variance equal to. S., with variance equal to. and
S.3, with variance equal to.5 (non-symmetric). S-Rule: For detecting stratification patterns be present in the sequence of the values of the statistic on the control chart we use one rule, which is based on the number of overlapping runs of length k in the center zone of the control chart (C.L-ó, C.L+ó). T: Trend patterns require a continuous movement to one direction of the chart according to a trend parameter Ä. T., with a trend parameter Ä=., T., with a trend parameter Ä=.5,and T.3, with a trend parameter Ä=.5 Ô-Rule: For detecting linear trend in the sequence of the values of the statistic on the control chart we use one rule, which is based on the number of the times when there is a positive difference in the sequence X i+w -X i. The following figures give the empirical power as well as the error rates of each rule when applied to each scheme. For example, in Figure we give the power (right detection rate) for the rule based on the number of overlapping runs of length k in the one side of the C.L. for detecting cyclical pattern. In this figure we give the power for values of k=,3,,5, while in Figure,,,,,,, Figure : Right Detection Rates of Model C. k= k= k= k=,,,, Figure 9: Numerical Error Rates of Model C. k= k=,5,,3 k=, k=, Figure : Right Detection Rates of Model C., Figure : Right Detection Rates of Model C.3, k=, k=,, k=, k= Figure 7: Numerical Error Rates of Model C. Figures through, correspond to the detection power and the error rate of the proposed rule for the models encoded as C., C. and C.3. The range of the parameter k is from to 5. Greater values of k have not any effect in the power of the rule. Moreover, Figures through 7, correspond to the detection power and the error rate for the models encoded as S., S. and S.3. The range of the parameter k is also from to 5. Greater values of k have not any effect in the power of the rule. Figure : Numerical Error Rates of Model C.3 Furthermore, Figures through 3, correspond to the detection power and the error rate for the models encoded as T., T. and T.3. In that figures we tried out the values 5, 7,, for the window parameter w.
,99,9, Figure : Right Detection Rates of Model S. k= k= IV. CONCLUSIONS Regarding the tests for detecting cycles in the process, we may denote that the rule has better performance when we choose the k parameter be equal to 3 or. Taking into consideration the numerical errors, we can see that all the schemes (C., C. and C.3) have the same behavior for all values of k. We may easily observe that for all values of k, all the three schemes have error rates in the range 5% to %. Thus, we may suggest the use of or k= in order to have a better performance when applying a test using the newly proposed rule for detecting cyclical behavior. Finally, we may also note that for the C.3 scheme, the Right Detection rate, for and k=, is about three times lower than in the other two schemes.,5, k= k=,,5,5 k=, k= Figure 3: Numerical Error Rates of Model S.,95,3,9 Figure : Right Detection Rates of Model S.3,5 k= k=,,,5 k=,75, k=,5 Figure : Right Detection Rates of Model S.,,5, k= k=, Figure 7: Numerical Error Rates of Model S.3,5,, w= w= Figure 5: Numerical Error Rates of Model S. In the following section we briefly discuss some properties of the proposed rules. Figure : Right Detection Rates of Model T.
,9,75,5,5 w= w= case, we propose the use w= or w= as well as n > 5. In Table IV, we present our proposals. Non-random pattern Test Statistic Values Cycles M m,k or k= Stratification M m,k k=, or k= Trend B w, w= or greater Table VI: Characterization Scheme Figure 9: Numerical Error Rates of Model T.,75,5,5,75,5 w=,5 w= w= w= Figure : Right Detection Rates of Model T.3, Figure : Right Detection Rates of Model T.,,, w= w=,,, w=, w= Figure 3: Numerical Error Rates of Model T.3 Figure : Numerical Error Rates of Model T. From the examination of the schemes encoded as S., S., and S.3, we may conclude that the power of the proposed test for that case depends on the scheme. For the T. scheme all the values are good for detecting (with high rates) stratification in the data, while, for the T. scheme the behavior is the same but the rates are considerably lower than in T.. Moreover, for T.3 scheme, rules based on all the values of k, except for, have a performance of % - 9% in detecting the nonrandom pattern. Taking into consideration the numerical error, which decreases as n increases, we suggest the use of or k=. Finally, for the detection of trend in the process, the test that we developed, is not sensitive enough in detecting trend with Ä=. or lower, even though for large values of n. Also, we observe that the power of the test, is an increasing function of n, Ä and w. Thus, we suggest that the proposed rule may be applied for detecting trend when the Ä is greater than.5. In that ACKNOWLEDGMENT The work of Sotiris Bersimis was sponsored by General Secretariat of Research and Technology of Greece under grant PENED. The work of Athanasios Rakitzis is supported by the State Scholarship Foundation of Greece. REFERENCES []. Bissel, A.F. (97). An Attempt to Unify the Theory of Quality Control Procedures, Bulletin in Applied Statistics, 5, 3-. []. Klein M. (). Two alternatives to the Shewhart X- bar Control Chart, Journal of Quality Technology, 3, 7-3 [3]. Khoo M.B.C. (3). Design of Runs Rules Schemes. Quality Engineering,,, 7-3. []. Montgomery, D. C. (). Introduction to Statistical Quality Control, New York: John Wiley [5]. Nelson, L.S. (9). The Shewhart Control Chart Test for Special Causes, Journal of Quality Technology,,, 337-39.
[]. Rakitzis A. (). Shewhart Control Charts with Stopping Rules Based on Runs. Unpublished MSc Thesis, University of Piraeus, Department of Statistics and Insurance Science, Greece. [7]. Roberts, S.W. (95). Properties of Control Chart Zone Tests, The Bell System Technical Journal, 37, 3-. []. Page, E.S. (955). Control Charts with Warning Lines, Biometrika,, 3-57. [9]. Western Electric Company (95). Statistical Quality Control Handbook, AT&T., Indianapolis, IN.