Monitoring feedback-controlled processes using adaptiveT2schemes

Kaibo Wang, Fugee TsungInternational Journal of Production Research 2007 · DOI

TL;DR

This paper introduces an "adaptive T² chart" to improve the monitoring of feedback-controlled industrial processes. By accounting for the dynamic, time-varying patterns that arise from feedback loops, this new method significantly enhances the ability to detect process shifts compared to traditional statistical process control (SPC) techniques.

What Problem Does This Paper Address?

Industrial processes often employ feedback control loops (Engineering Process Control - EPC) to maintain output at a target level. While effective for immediate adjustments, these loops introduce dynamic patterns in the process output, especially when faults or shifts occur. Traditional Statistical Process Control (SPC) charts, designed for static or constant shifts, struggle to detect these dynamic shifts effectively. This leads to a "window of opportunity" problem, where faults are missed, and the process can drift without detection. The paper aims to bridge this gap by developing an SPC method that explicitly considers and leverages these dynamic shift patterns.

Methods

The core of the paper is the development and evaluation of an "adaptive T² chart." This chart is an extension of the traditional T² chart, which is used for multivariate process monitoring. The key innovation lies in its ability to adapt to time-varying shift directions.

Here's a breakdown of the methods:

  • Modeling Dynamic Shifts: The paper first formulates and analyzes the dynamic shift patterns observed in processes controlled by Proportional Integral (PI) and Minimum Mean Square Error (MMSE) controllers. This involves understanding how these controllers influence the process output and control actions when a shift occurs.
  • Adaptive T² Chart: The proposed adaptive T² chart is based on the Uniformly Most Powerful (UMP) unbiased test. Instead of assuming a fixed shift direction, it uses a forecasting algorithm to estimate the time-varying shift direction at each step. This allows the chart to be sensitive to the evolving nature of process shifts.
  • Forecasting Algorithm: A model-free Exponentially Weighted Moving Average (EWMA) forecasting algorithm is employed to estimate the dynamic shift direction. For processes exhibiting strong oscillations, a modified "Oscillated EWMA" (OEWMA) is also proposed to improve forecasting accuracy.
  • Simulation and Evaluation: The performance of the adaptive T² chart is evaluated through Monte Carlo simulations. It is compared against existing methods like the U₀ and U∞ charts, as well as the standard T² and multivariate EWMA (MEWMA) charts.

Datasets

Not discussed in the paper. The evaluation relies on simulated data generated from models of PI and MMSE controlled processes.

Key Insights and Ideas

  • Dynamic Shifts are Informative: The paper strongly emphasizes that the dynamic patterns generated by feedback controllers in response to shifts are not just noise but contain crucial information for fault detection. Ignoring this information significantly hinders SPC performance.
  • Adaptability is Key: Traditional SPC charts are too rigid. The adaptive T² chart's ability to dynamically adjust its monitoring direction based on estimated shift trends is its primary strength, allowing it to capture shifts that static charts would miss.
  • EWMA/OEWMA for Forecasting: The use of EWMA (and its extension, OEWMA) as a model-free forecasting method for shift direction is a practical approach. It avoids the complexities of explicit time-series model fitting and proves effective, especially OEWMA for oscillating processes.
  • Trade-offs in Smoothing: The smoothing parameter (λ) in EWMA/OEWMA presents a trade-off. A higher λ leads to faster tracking of shifts but makes the forecast more susceptible to noise. A lower λ provides more robustness but might slow down the detection of large shifts. This allows for flexible tuning based on specific needs.

Key Results

  • The proposed adaptive T² chart demonstrates significantly improved detection sensitivity over intended shift ranges compared to existing methods.
  • For small shifts, the adaptive charts (AT2-E and AT2-OE) outperform the standard T² and U₀ charts.
  • While MEWMA and U∞ charts also perform well for small shifts, their large shift performance deteriorates significantly, whereas the adaptive T² charts maintain good performance across a range of shift magnitudes.
  • The AT2-E chart shows a slight edge for small shifts, while the AT2-OE chart is more robust to strong oscillations and has a higher probability of detecting shifts with short runs, making it recommended for general situations.

Limitations and Future Work

  • Determination of Control Limits: The paper mentions that determining the control limit (h₃) for the adaptive T² chart involves numerical methods and can be computationally intensive. While a software program is provided, further research into analytical methods for control limit calculation is suggested.
  • ARL Calculation: The study relies on Monte Carlo simulations for ARL computation. Developing more efficient analytical methods for ARL calculation for the adaptive T² chart is an area for future work.
  • Model Complexity: While the paper focuses on PI and MMSE controllers, extending the adaptive T² approach to more complex feedback control systems could be a valuable future direction.
  • Real-world Implementation: The paper presents simulation results. Further validation with real industrial data would be beneficial to confirm the practical effectiveness of the proposed method.

Summary updated

Abstract

The integration of engineering process control (EPC) and statistical process control (SPC) has drawn wide attention because of its capability to minimize long-term process variation and improve product quality. However, dynamic shift patterns are usually observed when feedback-control loops exist, which consequently lead to poor charting performance. In this paper, the dynamic patterns in mean shifts of proportional integral (PI) controlled and minimum mean square error (MMSE) controlled processes are formulated and analysed. In conjunction with a model-free forecasting algorithm, an adaptive T 2 charting procedure is applied to improve detection power. This adaptive procedure is constructed upon the uniformly most powerful (UMP) unbiased test with the consideration of time-varying patterns. Monte Carlo simulations have shown that the proposed strategy possesses significantly improved detection sensitivity over intended shift ranges.
BibTeX
@article{wang2007monitoring,
  author = {Kaibo Wang and Fugee Tsung},
  title = {{Monitoring feedback-controlled processes using adaptive<b><i>T</i></b><sup>2</sup>schemes}},
  journal = {International Journal of Production Research},
  year = {2007},
  doi = {10.1080/00207540701325488},
}

Loading related content...

Figure