Optimal designs of the side sensitive synthetic chart for the coefficient of variation based on the median run length and expected median run length

The side sensitive synthetic chart was proposed to improve the performance of the synthetic chart to monitor shifts in the coefficient of variation (γ), by incorporating the side sensitivity feature where successive non-conforming samples must fall on the same side of the control limits. The existing side sensitive synthetic- γ chart is only evaluated in terms of the average run length (ARL) and expected average run length (EARL). However, the run length distribution is skewed to the right, hence the actual performance of the chart may be frequently different from what is shown by the ARL and EARL. This paper evaluates the entire run length distribution by studying the percentiles of the run length distribution. It is shown that false alarms frequently happen much earlier than the in-control ARL (ARL0), and small shifts are often detected earlier compared to the ARL1. Subsequently, this paper proposes an alternative design based on the median run length (MRL) and expected median run length (EMRL). The optimal design based on the MRL shows smaller out-of-control MRL (MRL1), which shows a quicker detection of the out-of-control condition, compared to the existing design, while the results from the optimal design based on the EMRL is similar to that of the existing designs. Comparisons with the synthetic-γ chart without side sensitivity shows that side sensitivity reduces the median number of samples required to detect a shift and reduces the variability in the run length. Finally, the proposed designs are implemented on an actual industrial example.


Introduction
Control charts are useful tools to detect the presence of assignable cause(s) which results in an out-of-control process. By convention, control charts monitor the process mean (μ) and/or standard deviation (σ), where shifts in μ and/or σ shows the presence of assignable cause(s). For example, readers can refer to the studies by Aslam et al. [1], Riaz et al. [2], Abujiya et al. [3], and many others. samples taken until a shift of a specific magnitude is detected. The ARL 1 requires the shift size to be specified. It is difficult to specify the shift size in a lot of practical scenarios, hence the side sensitive synthetic-γ chart is also evaluated in terms of the EARL, which is the expected average number of samples taken to detect a shift that is specified as a range of values. Evaluating a chart's performance based solely on the ARL might lead to confusion on the actual performance of the chart [25]. This is because, for in-control processes or out-of-control processes with small shift sizes, the run length distribution is highly skewed to the right [26]. For a right-skewed distribution, the median is smaller than the mean, hence the median run length (MRL) will be smaller than the ARL. When the in-control MRL (MRL 0 ) is smaller than the ARL 0 , more than 50% of the in-control run lengths will be smaller than the ARL 0 . In other words, for more than 50% of the time, the false alarm will happen before what is indicated by the ARL 0 . When this happens, practitioners would have reduced confidence towards the ARL 0 as a performance measure. Hence, in this paper, the performance of the side sensitive synthetic-γ chart is evaluated based on its' entire run length distribution to have a clearer idea of its' actual performance. An analysis of the in-control and out-of-control percentiles (for known shift sizes) and expected percentiles (for unknown shift sizes) will be conducted.
The optimal charting parameters of the existing side sensitive synthetic-γ chart is obtained based on two designs, where the first design optimizes the ARL 1 (for known shift sizes) while the second optimizes the EARL (for unknown shift size). For both designs, constraints in ARL 0 needs to be satisfied. However, since the run length distribution is skewed to the right (for incontrol and out-of-control run lengths for small shift sizes), alternative designs are proposed to obtain the optimal charting parameters of the side sensitive synthetic-γ chart. In this paper, alternative designs are proposed where the optimal charting parameters are obtained to minimize the out-of-control MRL (MRL 1 ) and the expected MRL (EMRL), subject to constraints in the MRL 0 . This is because the MRL is a more accurate measure of performance when the run length is skewed.
Designs based on the MRL are available for several charts in the literature, among some of the more recent ones are studies on the MRL performance of the synthetic � X chart by Hu et al. [27], optimal designs based on the MRL and/or EMRL for the one-sided exponential CUSUM chart [28], one-sided exponential EWMA chart [29], EWMA-γ chart [30], EWMA � X chart [31], variable sample size � X chart [32], synthetic np chart [33], multivariate synthetic |S| chart [34], double sampling � X chart [35], and many others. However, the design of the side sensitive synthetic-γ chart is not available in the literature. This paper will fill this gap.
The rest of the paper is organized as follows. The next section gives an overview of the side sensitive synthetic-γ chart, as well as the formulae to evaluate the ARL, standard deviation of the run length (SDRL), EARL and the percentiles of the run length distribution. Section 3 analyses the percentiles of the run length distribution by adopting the optimal charting parameters by Yeong et al. [24]. Subsequently, Section 4 proposes the algorithms to obtain the optimal charting parameters based on the MRL 1 and EMRL, and illustrates the optimal charting parameters, MRL 1 and EMRL based on several numerical examples. This is followed by a comparison with the synthetic-γ chart without side sensitivity in Section 5, and the implementation of the proposed designs on an actual industrial example in Section 6. Finally, some concluding remarks are given in Section 7.

Side sensitive synthetic-γ chart
The synthetic-γ chart works by waiting until the second sample to fall outside the control limits before deciding whether the process is in-control or out-of-control. For the synthetic-γ chart without the side sensitivity feature, when the sample coefficient of variation ðĝÞ falls within the UCL and LCL, the sample is a conforming sample, while ifĝ > UCL orĝ < LCL, the sample is a non-conforming sample. The number of conforming samples between two successive non-conforming samples (including the ending non-conforming sample) is referred to as the conforming run length (CRL). For example, when there are three conforming samples between two successive non-conforming samples, the CRL = 4. When CRL�L, where L is a pre-determined threshold set by the practitioner, the synthetic-γ chart will produce an out-ofcontrol signal. To give the synthetic-γ chart a head-start, the first CRL counts the number of conforming samples until the first non-conforming sample appears.
The main difference between the side sensitive synthetic-γ chart and the synthetic-γ chart is that successive non-conforming samples must fall on the same side of the control limits. For example, by referring to the CRL sub-chart of the side sensitive synthetic-γ chart in Fig 1, since the first non-conforming sample (Sample 2) falls above the UCL, only samples that fall above the UCL are considered to be non-conforming samples. Although Sample 5 falls outside the control limits, it is still a conforming sample as it falls below the LCL. The next non-conforming only occurs in Sample 7. Hence, the CRL = 5.
The LCL and UCL of the side sensitive synthetic-γ chart are computed as where K represents the control limit coefficient, while m 0 ðĝÞ and s 0 ðĝÞ are the in-control mean

PLOS ONE
Side sensitive synthetic coefficient of variation chart based on median and expected median run lengths and standard deviation ofĝ, respectively. Although m 0 ðĝÞ and s 0 ðĝÞ have no closed forms, the following approximations by Reh and Scheffler [36]will be used, i.e.
and s 0 ðĝÞ � g 0 ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffiffi ffi ffi ffi ffi ffi ffi ffi ffi ffi 1 n where γ 0 and n are the in-control γ and sample size, respectively. A Markov chain approach is adopted to evaluate the ARL, SDRL, EARL and the percentiles of the run length distribution. The formulae to evaluate the ARL, SDRL and EARL is shown in Yeong et al. [24] and is reproduced in S1 Appendix for ease of reference.
To study the percentiles of the run length distribution, the probability mass function (p.m. f.) and cumulative distribution function (c.d.f.) of the run length needs to be obtained. The p. m.f. and c.d.f. of the run length for the side sensitive synthetic-γ chart can be computed as [37] and where l2{1,2,3,. . .}, q and 1 are as defined in Equations (A5) and (A6), while Q and r are as defined in Equation (A1). The (100θ) th percentile of the run length distribution of the side sensitive synthetic-γ chart can be obtained from Eq (6) by searching for l θ such that [26] PðRL � l y À 1Þ � y and PðRL � l y Þ > y; where 0<θ<1. For example, the MRL, which is the 50 th percentile of the run length, can be obtained from Eq (7) by setting θ = 0.5. The in-control percentiles can be obtained by setting γ = γ 0 when computing the transition probabilities of Q through Equations (A2) to (A4), while the out-of-control percentiles are obtained by setting γ = τγ 0 , where τ is the shift size. The percentiles of the run length can only be obtained if τ can be known in advance. However, practitioners usually find it difficult to specify τ due to insufficient data. Furthermore, τ may not be deterministic and may vary according to some unknown stochastic model [6]. Under such conditions, the expected percentiles of the run-length distribution, E(l θ ) are evaluated instead. E(l θ ) does not require τ to be specified as a specific value. Instead, τ only needs to be specified as a range of possible values, (τ min ,τ max ). E(l θ ) can be computed as follows: where f τ (τ) is the probability density function (p.d.f.) of τ. It is assumed that f τ (τ) is a uniformly distributed continuous random variable over the interval (τ min ,τ max ) [6]. The integral in Eq (8) is approximated using the Gauss-Legendre quadrature [38].

Analysis of the percentiles of the run length distribution
In this section, the percentiles of the run length distribution are analysed based on the optimal charting parameters by Yeong et al. [24]. Yeong et al. [24] adopted two types of designs. In the first design, the optimal charting parameters are obtained to minimize ARL 1 , subject to constraints in the ARL 0 , while the second design obtains the optimal charting parameters to minimize the EARL, subject to the same constraints in ARL 0 . For ease of reference, the first and second designs are referred to as ARL-based and EARL-based designs, respectively. Both the in-control and out-of-control run length percentiles are analysed in this section. Table 1 shows the optimal charting parameters (L, LCL and UCL), the ARL 1 and SDRL 1 of the ARL-based design of the side sensitive synthetic-γ chart for γ 0 = 0.05,n2{5,7,10,15} and τ2 {1.1,1.2,1.3,1.5,2.0}. Note that τ>1 indicates an upward shift in γ 0 , while τ<1 indicates a downward shift. Practitioners are often more interested in detecting an upward shift, since an upward shift results in a larger s m ratio which shows that there is increased variability in the process due to a relatively large σ compared to μ. The increased variability is usually due to the presence of assignable cause(s). It is important to detect the presence of such assignable cause (s) so that they can be removed to reduce variability in the process. Comparatively, detecting a downward shift is less crucial as a downward shift shows decreased variability. Thus, this paper only considers τ>1. The optimal charting parameters in Table 1 are obtained to minimize the ARL 1 , subject to ARL 0 = 370.4. For example, for n = 5 and τ = 1.1, adopting the charting parameters (L,LCL, UCL) = (42,0.0017,0.0924) will result in the smallest ARL 1 , while maintaining the ARL 0 as 370.4. The (ARL 1 ,SDRL 1 ) = 64.74,84.69) when these optimal charting parameters are adopted.
To get a better picture on the actual run length performance of the ARL-based side sensitive synthetic-γ chart, the entire in-control and out-of-control run length distribution is analyzed. Table 2 shows the 5 th until 95 th in-control percentiles of the side sensitive synthetic-γ chart, while Table 3 shows the corresponding out-of-control percentiles. The percentiles are obtained by substituting the optimal charting parameters in Table 1, together with the corresponding γ 0 , n and τ, into Eq (7).
Analysing the percentiles of the run length distribution will give us a better idea on the actual run length performance of the chart. For example, by referring to Table 2, when the optimal charting parameters for the ARL-based design is adopted, the MRL 0 obtained is between 211 to 251, with smaller values obtained for smaller τ and n. This is significantly different from the ARL 0 of 370.4, which shows that for 50% of the time, the false alarm will happen much earlier than the 370 th sample. When practitioners design the chart based on an ARL 0 of 370.4, they would expect than most of the time, false alarms only happen by the 370 th sample. However, this is not the case from the MRL 0 values obtained in Table 2. In fact, 370 falls between the 60 th and 70 th percentile. This shows the in-control run length distribution is skewed to the right, and interpreting the frequency of false alarms based purely on the ARL 0 is not accurate. Furthermore, studying the difference between the extreme percentiles (for example, the difference between the 5 th and 95 th percentile) enables practitioners to have a better idea on the variation of the run length. A similar trend is observed for the out-of-control percentiles, especially for small values of τ. In particular, there is a significant difference between the ARL 1 in Table 1 with the MRL 1 in Table 3. For example, for n = 5 and τ = 1.1, Table 3 shows an MRL 1 of 29, but the ARL 1 is 64.74 from Table 1. This shows that for more than 50% of the time, the out-of-control condition will be detected before what is indicated by the ARL 1 . In most cases, the ARL 1 lies close to the 70 th percentile. Hence, evaluating the side sensitive synthetic-γ chart based purely on the ARL 1 will result in an inaccurate representation of the actual performance of the chart when it is implemented since in most cases the shift is detected earlier than the ARL 1 .
From Table 3, the difference between the extreme percentiles reduces as τ and n increases. This shows that there is less variation in the out-of-control run lengths for larger τ and n. For example, for n = 5 and τ = 1.1, the difference between the 5 th and 95 th percentile is 237, while for n = 5 and τ = 2.0, the corresponding difference is only 3. Similarly, for n = 15 and τ = 1.1, the difference between the extreme percentiles is 113 (compared to the corresponding difference of 237 for n = 5 and τ = 1.1).
The percentiles in Tables 2 and 3 can only be obtained if τ can be specified in advance. Since τ cannot be specified in certain practical scenarios, this section also analyses the expected percentiles. In this paper, (τ min ,τ max ) is set as (1,2]. Table 4 shows the optimal charting parameters (L, LCL and UCL) and the EARL of the EARL-based design of the side sensitive synthetic-γ chart for γ 0 = 0.05 and n2{5,7,10,15}. Tables 5 and 6 show the in-control percentiles and out- Table 3. Out-of-control percentiles of the run length distribution for the ARL-based side sensitive synthetic-γ chart for γ 0 = 0.05, n2{5,7,10,15} and τ2  {1.1,1.2,1.3,1.5

PLOS ONE
Side sensitive synthetic coefficient of variation chart based on median and expected median run lengths of-control expected percentiles, respectively, when the charting parameters in Table 4 are adopted. The expected percentiles are obtained by substituting the optimal charting parameters in Table 4, together with the corresponding γ 0 , n and (τ min ,τ max ) = (1,2], into Eq (8).
The in-control percentiles in Table 5 show a similar trend as the ARL-based design. The MRL 0 ranges from 219 to 226, which is significantly smaller than the ARL 0 of 370.4. From Table 6, the EMRL is also significantly smaller than the EARL 1 . For example, for n = 5, the EMRL is 9.02, while the corresponding EARL is 16.90 from Table 4. The expected percentiles in Table 6 decreases for larger n. Similarly, the difference between the extreme expected percentiles decreases for larger n, which shows less variation in the expected percentiles for larger n.

MRL and EMRL-based design of the side sensitive synthetic-γ chart
As the MRL and EMRL provide more accurate results than the ARL and EARL when the run length distribution is skewed, this section proposes an alternative design where the optimal charting parameters which minimize the MRL 1 and EMRL are obtained, subject to constraints in the MRL 0 .
The following are the algorithms to obtain the optimal charting parameters based on the MRL-based design.

By using the (LCL,UCL) in
Step 3, numerically search for l 0.5 that satisfies Eq (7) for the γ 0 , n and τ determined in Step 1. The MRL 1 = l 0.5 .
6. Repeat Steps 3 to 5 until the MRL 1 for L+1 is larger than the MRL 1 for L. This (L,LCL,UCL) combination will be the optimal charting parameters for the MRL-based side sensitive synthetic-γ chart. If there is more than one combination of (L,LCL,UCL) with the smallest MRL 1 , the combination with the smallest value for the difference between l 0.95 and l 0.05 is chosen as the optimal charting parameters, where l 0.05 and l 0.95 are the 5 th and 95 th percentiles of the out-of-control run length distribution, respectively.
To ensure a fair comparison between the ARL-based and MRL-based design, ξ is set to be equivalent to the MRL 0 in Table 2. For example, by referring to the MRL 0 for n = 5 and τ = 1.1 in Table 2, ξ = 211. Table 7 shows the optimal charting parameters for the MRL-based design, and the 5 th percentile (Q05), MRL 1 , 95 th percentile (Q95), ARL 1 and ARL 0 when these optimal charting parameters are adopted are also shown. By comparing the optimal charting parameters in Table 1 with that in Table 7, it can be observed that the optimal L for the ARL-based design is generally larger for small values of τ and n compared to the MRL-based design. For example, for n = 5 and τ = 1.1, the optimal L = 42 for the ARL-based design, while the optimal L = 22 for the MRL-based design. The smaller optimal L for the MRL-based design is also associated with a smaller conforming region, as shown by larger values of LCL and smaller values of UCL. For example, for n = 5 and τ = 1.1, the optimal (LCL, UCL) = (0.0017, 0.0924) for the ARL-based design, while the optimal (LCL, UCL) = (0.0043, 0.0898) for the MRL -based design.
Next, comparing the MRL 1 in Table 3 with that in Table 7 shows that the MRL-based design results in smaller MRL 1 compared with that in the ARL-based design, especially for small values of τ. For example, for n = 5 and τ = 1.1, the MRL 1 = 29 for the ARL-based design in Table 3, while the MRL 1 = 22 for the ARL-based design in Table 7. This shows that the MRLbased design results in better MRL 1 performance compared with the ARL-based design. Note that both designs have the same MRL 0 .
Both the MRL-based and ARL-based designs show similar ARL 1 . For example, for n = 5 and τ = 1.1, the ARL 1 = 64.74 for the ARL-based design in Table 1, while the ARL 1 = 63.38 for the MRL-based design in Table 7. This shows that the MRL-based design results in a smaller median number of samples to detect the shift, but with a similar average number of samples to detect the shift.
The ARL 0 is set as 370.4 for the ARL-based design, while the ARL 0 for the MRL-based design in Table 7 is between 344.30 to 375.27. Note that the MRL-based design does not fix the value for ARL 0 . The ARL 0 for the MRL-based design is generally smaller than that of the ARLbased design, however, the difference is not large.
The MRL-based design in Table 7 can only be implemented if τ can be specified in advance. This paper also considers the EMRL-based design for cases where τ is unknown. Similar steps as shown in paragraph 2 for the MRL-based design are adopted, but Steps 1, 4 and 6 are replaced with the following.
6. Repeat Steps 3 to 5 until the EMRL for L+1 is larger than the EMRL for L. This (L,LCL, UCL) combination will be the optimal charting parameters for the EMRL-based side sensitive synthetic-γ chart.
Similar to the EARL-based design, (τ min ,τ max ) is set as (1,2]. To ensure a fair comparison between the EARL-based and EMRL-based design, ξ is set to be equivalent to the MRL 0 in Table 5. Table 8 shows the optimal charting parameters for n2{5,7,10,15} and its Table 7. Optimal charting parameters and the corresponding Q05, MRL 1 , Q95, ARL 1 and ARL 0 for the MRL-based side sensitive synthetic-γ chart for γ 0 = 0.05, n2  {5,7,10,15} and τ2{1.1,1.2,1.3,1.5,2 corresponding 5 th expected percentile (EQ05), EMRL, 95 th expected percentile (EQ95), EARL and ARL 0 when these optimal charting parameters are adopted. By comparing the optimal charting parameters for the EMRL-based design in Table 8 with the corresponding optimal charting parameters for the EARL-based design in Table 4, it can be observed that the optimal L in Table 8 is smaller than the optimal L in Table 4. Furthermore, the conforming region for the EMRL-based design is smaller than that of the EARL-based design, as shown by the larger LCL and smaller UCL in Table 8, compared to the LCL and UCL in Table 4. Minimal differences are shown between the EQ05, EMRL and EQ95 values of the EARL and EMRL-based designs, by comparing Tables 6 and 8. The EARL values in Tables 4  and 8 are also similar. Hence, the EARL and EMRL-based designs show similar performance. The ARL 0 of the EMRL-based design is slightly smaller than the ARL 0 of the EARL-based design, but the difference is not very large. Note that the ARL 0 of the EARL-based design is fixed as 370.4.

Comparisons
In this section, the MRL and EMRL-based designs of the side sensitive synthetic-γ chart is compared with the corresponding designs for the synthetic-γchart without side sensitivity. To obtain the MRL and EMRL-based designs for the synthetic-γ chart without side sensitivity, a similar procedure as shown in Section 4 is adopted, but modified for the synthetic-γ chart without side sensitivity. Table 9 shows the Q05, MRL 1 and Q95 for the MRL-based designs of these two charts, while Table 10 shows the EQ05, EMRL 1 and EQ95 for the EMRL-based designs.
From Table 9, the side sensitive synthetic-γ chart shows smaller MRL 1 and Q95 compared to the synthetic-γ chart without side sensitivity, especially for small values of τ. For example, for n = 5 and τ = 1.1, (MRL 1 ,Q95) = (32,343) for the synthetic-γ chart without side sensitivity, whereas (MRL 1 ,Q95) = (22,231) for the side sensitive synthetic-γ chart. This shows that incorporating side sensitivity reduces the median number of samples required to detect the shift, and at the same time reduces the variability in the run length due to a smaller difference between Q95 and Q05. This is consistent with the results by Yeong et al. (2021), which shows that the side sensitive synthetic-γ chart shows better ARL performance than the synthetic-γ chart without side sensitivity. From Table 10, a similar conclusion is reached for the EMRLbased design. Thus, it can be concluded that the side sensitive synthetic-γ chart shows better performance than the synthetic-γ chart without side sensitivity.

Illustrative example
This section shows the implementation of the MRL and EMRL-based design on an actual industrial example. The example was also adopted by Yeong et al. [24] who proposed the side sensitive synthetic-γ chart. The example is from a sintering process where compressed metal powder is heated to a temperature that allows bonding of the individual particles. The strength of the bond between particles is influenced by pore shrinkage [39]. One of the characteristics that is related to pore shrinkage is the pressure test drop time (T pd ) from 2 bar to 1.5 bar, which must be larger than 30 seconds.
Sintering steel with a heterogeneous microstructure and an irregular grain size will lead to an anomalous increase in the standard deviation of T pd (σ pd ), which will result in a change in the correlation structure between the mean of T pd (μ pd ) and σ pd and subsequently results in shifts in the coefficient of variation of T pd (γ pd ). Hence, the special cause can be detected by monitoring γ pd . Furthermore, Castagliola et al. [6] through a regression study showed that σ pd = γ pd ×μ pd , which provides additional evidence that the process can be monitored by detecting changes in γ pd . Table 11 (left) shows a Phase I dataset of m = 20 samples, each with a sample size n = 5. � X k ; S k andĝ k denotes the sample mean, standard deviation and coefficient of variation, Table 9. Comparison of the Q05, MRL 1 and Q95 of the MRL-based synthetic-γ chart (without side sensitivity) and side sensitive synthetic-γ chart for γ 0 = 0.05, n2{5,7,10,15} and τ2{1.1,1.2,1.3,1.5 respectively, of the k th sample, k = 1, 2, . . . 20. The data is taken from Castagliola et al. [6]. Castagliola et al. [6] showed that the Phase I data were in-control, with an estimated in-control γ ðĝ 0 Þ of 0.417 based on a root-mean-square computation. Table 11 (right) shows the Phase II data after the occurrence of a special cause that increased process variability. A shift of 25% in the coefficient of variation shows that something is wrong in the production of the parts. Hence, the side-sensitive synthetic-γ chart is designed to detect a shift of τ = 1.25. The MRL-based design in Section 4 is adopted to obtain the optimal charting parameters, where ξ = 250. The optimal charting parameters are (L,LCL, UCL) = (7,0,0.8418), which results in (Q05,MRL 1 ,Q95) = (1,7,76). By comparison, the ARLbased design by Yeong et al. [24] results in (Q05,MRL 1 ,Q95) = (1,10,68). Fig 2 shows the γ sub-chart of the side-sensitive synthetic-γ chart by adopting the MRL-based design.
From Fig 2, two non-conforming samples are detected, i.e. Samples 3 and 7, with CRL of 3 and 4, respectively. Since both CRLs are less than 7, out-of-control signals are produced at Samples 3 and 7. This shows that the side-sensitive synthetic-γ chart can show a quick detection of the special cause. The ARL-based design by Yeong et al. [24] also detected the out-ofcontrol condition at Samples 3 and 7.
The MRL-based design can only be adopted if τ can be specified. As the exact size of shift caused by the special cause may not be known, this section also considers the EMRL-based design with (τ min ,τ max ) = (1,2]. From the EMRL-based design in Section 4, the optimal charting parameters are (L,LCL,UCL) = (8,0,0.8504), which results in (EQ05,EMRL 1 ,EQ95) = (1.12,12.43,74.94). Note that as in the MRL-based design, ξ = 250. Fig 3 shows the γ sub-chart of the side-sensitive synthetic-γ chart by adopting the EMRL-based design. Similar to the MRL-based design, Samples 3 and 7 are detected as out-of-control samples.

Conclusions
This paper evaluates the performance of the side-sensitive synthetic-γ chart by studying its' entire run length distribution. An analysis of the run length distribution shows that false alarms frequently happens much earlier than that indicated by the ARL 0 . In addition, small shifts are often detected earlier compared to the ARL 1 . This will reduce the confidence towards the chart when the actual performance is frequently different from that indicated by the ARL.  Hence, the performance of the side-sensitive synthetic-γ chart should not be evaluated only in terms of the ARL, but should be accompanied by the entire run length distribution. This paper also proposes alternative designs for the side-sensitive synthetic-γ chart based on the MRL and EMRL. Compared to designs based on the ARL, the proposed design based on the MRL shows better MRL 1 performance, i.e. it requires a smaller median number of samples to detect shifts in γ. For the proposed design based on the EMRL, the performance is similar to that based on the EARL. Comparisons with the synthetic-γ chart without side sensitivity shows that side sensitivity reduces the median number of samples required to detect a shift and reduces the variability in the run length. Implementation of the proposed designs on an actual industrial example shows that they are efficient in detecting out-of-control conditions.