Modelling and diagnosis of feedback-controlled processes using dynamic PCA and neural networks

Dongfeng Shi, Fugee TsungInternational Journal of Production Research 2003 · DOI

TL;DR

This paper introduces a novel approach to diagnose faults in manufacturing processes that use feedback control. By combining dynamic Principal Component Analysis (PCA) with neural networks, the method effectively extracts key features from process data and accurately identifies the root cause of disturbances, outperforming traditional methods.

What Problem Does This Paper Address?

Manufacturing industries constantly strive for high-quality products and reduced variability. While Statistical Process Control (SPC) and Engineering Process Control (EPC) are established methods for achieving this, monitoring and diagnosing feedback-controlled processes presents a unique challenge. These systems inherently involve correlations between control actions and process outputs due to feedback mechanisms. Traditional SPC methods, which often assume independent data, struggle to effectively model and diagnose these complex, dynamic relationships. The paper aims to bridge this gap by developing a scheme that can accurately model these feedback-controlled processes and pinpoint the source of faults.

Methods

The core of the proposed solution lies in a two-pronged approach:

  • Dynamic Principal Component Analysis (DPCA): This technique is used for feature extraction. Unlike static PCA, DPCA considers time-shifted data (lagged observations) to capture the dynamic relationships between process output and control actions. This allows it to extract meaningful features that represent the "scatter ellipse" of the process data, including the length and inclination of its major axis. These extracted features are designed to be less sensitive to noise and redundancy in the original data.
  • Neural Networks (specifically Multilayer Perceptrons - MLPs): These are employed as classifiers. After DPCA extracts the relevant features, MLPs are trained to model the relationship between these features and the parameters of process disturbances. The paper highlights the advantages of MLPs, such as their data-driven nature, ability to approximate complex functions, and non-linear modeling capabilities. A conjugate gradient algorithm is used for efficient training.

Datasets

The paper does not use pre-existing datasets. Instead, it relies on simulated data generated to represent various fault scenarios in a feedback-controlled process. These simulations are based on a typical ARMA(1,1) model for disturbance and a PI controller. Four types of faults are simulated: shifts in disturbance parameters ($\phi$ and $\theta$), shifts in the mean of the disturbance, and changes in the variance of the disturbance. These simulated datasets are used for both training the neural networks and testing the performance of the proposed scheme.

Key Insights and Ideas

  • Dynamic relationships are crucial: The paper strongly emphasizes that ignoring the dynamic correlations in feedback-controlled processes leads to ineffective monitoring. DPCA is presented as a way to explicitly capture these dynamics.
  • Feature extraction is key to robust diagnosis: By using DPCA to extract features like the "scatter ellipse" characteristics, the method becomes more resilient to noise and reduces the dimensionality of the problem, which is beneficial for neural network training.
  • Neural networks excel at mapping complex relationships: The study demonstrates that MLPs are well-suited to learn the intricate mapping between the extracted DPCA features and the underlying disturbance parameters, enabling accurate fault diagnosis.
  • Mean and variance shifts are easier to detect: The simulation results indicate that changes in the mean and variance of the disturbance have a more pronounced impact on the process scatter, making them easier to diagnose compared to shifts in the ARMA parameters ($\phi$ and $\theta$).

Key Results

The proposed scheme, integrating DPCA and neural networks, demonstrated superior performance compared to a traditional minimal distance classifier.

  • Diagnosis Accuracy: The neural network-based approach achieved significantly higher diagnosis accuracy across various fault types and time lags (L). For instance, with L=2, the neural network achieved accuracies above 97% for parameter shifts and 98% for mean/variance shifts, while the minimal distance classifier was in the 75-91% range.
  • Impact of Time Lag (L): Increasing the time lag (L) in DPCA consistently improved the diagnosis accuracy for both methods. This is attributed to DPCA capturing more dynamic information with higher lags.
  • Robustness to Data Volume: The accuracy of the neural network approach remained high (above 85%) even with a reduced number of observations, indicating good generalization capabilities.

Limitations and Future Work

  • Simulation-based: The study relies on simulated data. While the ARMA(1,1) model and PI controller are representative, real-world processes might exhibit more complex dynamics.
  • Specific Fault Types: The study focuses on four specific types of disturbance parameter shifts. Future work could explore a wider range of fault types.
  • Tuning of L: The paper acknowledges that selecting the optimal time lag (L) for DPCA is critical and mentions existing criteria, but further research could refine this selection process.
  • Real-world validation: While the simulation results are promising, validation on actual industrial data would be a crucial next step to confirm the scheme's practical applicability.

Summary updated

Abstract

Statistical Process Control (SPC) and Engineering Process Control (EPC) have evolved rapidly in manufacturing industry. Typical feedback controllers, such as proportional integral derivative (PID) and exponentially weighted moving average (EWMA) controllers, are designed to maintain desirable operations by compensating for the effects of disturbances and changes in industrial processes according to process data. However, little attention is paid to the modelling and diagnosis of feedback-controlled processes. A new scheme to diagnose the root cause of faults in a feedback-controlled process is presented here by integrating dynamic principal component analysis (PCA) and neural networks. The dynamic PCA is introduced first to extract the dynamic relationship between control action and process output and to generate feature sets. Furthermore, the neural networks' classifier is trained by a scale-conjugated gradient algorithm using training samples. Simulation results show that the above scheme could model the relationship between the features extracted by dynamic PCA and disturbance parameters, and could improve the ability to diagnose the root causes of faults in comparison with traditional classifiers. In addition, this scheme offers a broad perspective in production variability reduction and optimization of the process controller.
BibTeX
@article{shi2003modelling,
  author = {Dongfeng Shi and Fugee Tsung},
  title = {{Modelling and diagnosis of feedback-controlled processes using dynamic PCA and neural networks}},
  journal = {International Journal of Production Research},
  year = {2003},
  doi = {10.1080/0020754021000020962},
}
Figure