An Improved Scheme for On-line Recognition of Control Chart Patterns
Authors: Adnan Hassan
Abstract
The key feature of control charts is the provision of the method to differentiate whether a particular process is operating within a statistically stable or an unstable state. Unstable processes may produce distinct time series patterns which are useful for diagnosis and troubleshooting. However, the point when a stable process actually starts to deteriorate and reaches recognizable patterns is normally unknown. It is important that a functional on-line scheme be developed to balance between unnecessary recognition of stable processes and premature recognition unstable processes. This paper proposes two alternative schemes for the on-line recognition of control chart patterns, namely: (i) a scheme based on direct continuous recognition, and (ii) a scheme based on recognition only when necessary. This study focuses on recognition of six control chart patterns plotted on the Shewhart X-bar chart, namely, random, shift-up, shift down, trend-up, trend-down, and cyclic. The artificial neural network (ANN) recognizer was used based on multilayer perceptrons (MLPs) architecture. The performance of the schemes were evaluated based on percentage correct recognition, average run length (ARL), type I and type II errors. The findings suggest that on-line recognition should be made only when necessary. Continuous recognition is not only wasteful, but also results in poorer results. The methodology proposed in this study is a step forward in realising a truly automated and intelligent SPC chart pattern recognition system.
Dear Hassan,
I have read your paper and found it very interesting. I have a query about the feature selection step of your proposed methods. You mentioned in section 2 that six features will be selected based on simulation studies and it is not clear why did you choose these specific features. How did you know that they are relevant to the problem at hand. Do you think using Fourier and Wavelet transformations to transfer the data streams from the time domain to the frequency domain can help in identifying the patterns.
Thank you,
Afify
Dear Afify,
Thank you for your interest in my work. The six features were selected from candidates of ten features, namely,
(a) Mean
(b) Mean square value (MSV)
(c) Median
(d) Standard deviation
(e) Range
(f) Cumulative sum (CUSUM)
(g) Skewness
(h) Kurtosis
(i) Slope
(j) Average autocorrelation
Since the data streams were time siries data (20 points), they could be easily summarized into the above features. However, our study revealed that, proper combination of the above features are very important to obtain good classification performance. We used fractional factorial design to identify the minimal set.
We have published a comparative study based on fully developed control chart patterns to compare the candidate features, see ref [8] and [9] for further details.
I agree with you that Fourier and Wavelet transformations could be other alternatives to transfer the data streams.
Thank you.
Adnan
Salaam Dr. Adnan.
Have you done any testing on other data set? Example wood data set.
Hope you still remember me. I’m back in UKM last year 2007.
Best wishes,
Shahnorbanun
Salaam Dr Shah,
Sorry for unable to reply earlier. Yes I still remember you. Good to know that you are back in UKM. I belief you are in the same dept with Dr Muriaty?
Regarding your question, we have not tested with other data set. Our research still focusing on the SPC data. FYI, we have also investigated on web-based CCPR and currently on multivatiate SPC.
Best Regards,
Adnan
Dear author,
Thank you very much for your contribution to IPROMS2008.
Could you please upload your video presentation similarly to what other authors have done in this session?
Kind regards,
Afify
Co-chair of the IDSS session