http://www.paper.edu.cn
0 INTRODUCTION
Statistical Process Control ( SPC ) has been widely used to monitor and improve the quality
and productivity of manufacturing processes. Most of research in SPC focuses on the charting
techniques. The practical applications of control charts now extend far beyond manufacturing
into biology, genetics, medicine, finance and other areas, see [1]. Current methods focus mostly
on monitoring and detecting constant ( step ) shifts in the process parameter. SPC methods for
detecting nonconstant ( or said ‘patterned’ ) shifts of the process parameter have not yet been
thoroughly studied.
With some patterned shifts exist, it is assumed that once a special cause initiates the
change of a process parameter away from its in-control ( IC ) value, the process parameter
continues to vary until a control chart detects such patterned changes, e.g., the process mean
of i observations after the beginning of the change is µ
0
+ θ
i−τ
, where τ is called ‘change-point’,
θ
i
represents a change pattern in the process mean per observation, and µ
0
is the mean when
the process is in-control. If the change θ
i
is a linear function, people always call it as ‘drifts’.
In real manufacturing process, patterned shifts are usually due to causes such as deterioration
of equipment, catalyst aging, waste accumulation or human causes.
Until now, most of the literatures concerned about the process with patterned changes are
mainly focused on the detection of process mean drifts. For example, [2], [3], [4], [5], [6], [7], [8]
and [9] have evaluated the performance of the Shewhart chart, CUSUM chart, EWMA chart,
Shewhart chart supplemented with runs rules and GLR chart when drifts in the process mean
exist, respectively. As refer to the other patterns of process mean shifts, [10] proposed a GLR
chart to monitor the residuals with an exponentially-patterned mean shifts in a time-series
model. [11] developed a reference-free Cuscore chart ( RFCuscore ) to monitor the same model
which can be implemented without knowing the change pattern {θ
i
}. However, SPC methods
for detecting other patterned mean shifts are quite few.
As we know, based on the generalized likelihood ratio approach for the on-line detection,
[12] proposed a CUSUM-like control chart called GLR which does not depend on reference
value. [13] investigated the similar control chart and their simulation results showed that this
chart had a robust performance for the magnitude of the shifts. However, such an idea has not
been further applied in detection of the patterned mean shifts. Therefore, in this paper, we
extend the generalized likelihood ratio method to the pattern shifts case and give a new control
chart ( called EML chart ). It will be shown that our proposed chart has the following good
features: 1. it can be easily designed and constructed because no additional parameter involved
except a smooth constant and an upper control limit; 2. By fully inheriting the advantages
of classical likelihood ratio tests, it is quite robust and sensitive to various types of patterned
mean shifts ( including the step shift and drifts ).
The rest of the paper is organized as follows: in the next section, we first give a brief
- 2 -