Diagnosis of Out-of-control Signals in Multivariate Statistical Process Control Based on Bagging and Decision Tree

Jing Jiang, Huaming SongAsian 2017 · DOI

Abstract

In this paper, we propose an ensemble method based on bagging and decision tree to resolve the problem of diagnosing out-of-control signals in multivariate statistical process control. To classify the out-of-control signals, we obtain a series of classifiers through ensemble learning on decision tree. Then we will integrate the classification results of multiple classifiers to determine the final classification. The experimental results show that our method could improve the accuracy of classification and is superior to other methods in terms of diagnosing out-of-control signals in multivariate statistical process control.

Introduction

In Multivariate Statistical Process Control (MSPC), managers face two primary hurdles: detection and diagnosis. While tools like Hotelling’s $T^2$ control charts are excellent at signaling that a process has deviated from its normal state, they are notoriously bad at explaining why. In a modern factory with dozens of interacting variables, simply knowing the process is "out of control" isn't enough. Engineers need to know which specific variable (or combination of variables) shifted so they can perform targeted maintenance. Previous attempts using Artificial Neural Networks (ANN) or Support Vector Machines (SVM) often struggled with accuracy or required immense computational resources. This paper seeks a more robust, accurate, and scalable way to classify these abnormal signals.

Methodology

The authors propose an ensemble method based on Bagging (Bootstrap Aggregating) and Decision Trees.

  1. Base Classifier: Standard classification trees are used. Decision trees are "unstable" learners, meaning small changes in the training data can lead to very different tree structures. While usually a weakness, this instability is a strength for ensemble learning.
  2. Bagging: To stabilize the predictions, the algorithm creates multiple new training sets by sampling the original data with replacement (bootstrapping).
  3. Ensemble Construction: A separate decision tree is trained on each bootstrapped set. The researchers used the Gini index to measure node impurity during tree growth and cost-complexity pruning to prevent overfitting.
  4. Majority Voting: When a new "out-of-control" signal appears, all trees in the ensemble vote on which variables caused the shift. The final diagnosis is determined by a simple majority vote.

Experimental Setup

The study utilized Monte Carlo simulations to generate three benchmark datasets based on real-world industrial scenarios:

  • Two-Variable Case: Derived from a lumber manufacturing process (measuring stiffness and bending strength).
  • Three-Variable Case: Based on a detergent-making process (measuring color, free oil percentage, and acidity).
  • Four-Variable Case: Based on ballistic missile testing data.

These datasets were standardized using zero-mean normalization to ensure that variables with different scales (e.g., temperature vs. pressure) didn't bias the model.

Key Results and Discussion

The Bagging-trees approach was compared against four competing methods: SAMME (a boosting algorithm), CART (a single decision tree), MLP (Multi-Layer Perceptron neural network), and MSCH (Multivariate Shewhart Chart).

Comparison of misclassification rates (%) across different methods, variable counts, and shift magnitudes.

As shown in the performance comparison, the Bagging-trees method achieved a total average misclassification rate of 9.83%, significantly outperforming SAMME (11.50%), CART (15.88%), MLP (29.71%), and MSCH (33.80%). The method's performance actually improves relative to other models as the number of variables increases. For the four-variable missile testing case, the Bagging-trees method maintained error rates as low as 1.13% for large shifts, while traditional methods like MSCH saw error rates climb above 60%.

Confusion matrices for the two-variable case, showing predicted vs. actual classes at different shift levels.

The diagnostic precision is further evidenced by the confusion matrices. The model treats diagnosis as a multi-class classification problem, capable of identifying $2^p - 1$ possible fault combinations. In the two-variable case, the model accurately identifies both individual variable shifts and joint shifts, with accuracy increasing alongside the magnitude of the shift ($\sigma$).

Conclusion

This research demonstrates that leveraging the inherent instability of decision trees through Bagging creates a powerful diagnostic tool for MSPC. By capturing a wider variety of "fault patterns," the ensemble approach provides superior accuracy and scalability compared to traditional neural networks and control charts. Future work will expand this framework to address other process abnormalities like downward shifts, trends, and cyclic variations, as well as validation on live industrial data streams.

Summary updated

Abstract

In this paper, we propose an ensemble method based on bagging and decision tree to resolve the problem of diagnosing out-of-control signals in multivariate statistical process control. To classify the out-of-control signals, we obtain a series of classifiers through ensemble learning on decision tree. Then we will integrate the classification results of multiple classifiers to determine the final classification. The experimental results show that our method could improve the accuracy of classification and is superior to other methods in terms of diagnosing out-of-control signals in multivariate statistical process control.
BibTeX
@inproceedings{jiang2017diagnosis,
  author = {Jing Jiang and Huaming Song},
  title = {{Diagnosis of Out-of-control Signals in Multivariate Statistical Process Control Based on Bagging and Decision Tree}},
  booktitle = {Asian},
  year = {2017},
  doi = {10.20849/abr.v2i2.147},
}
Figure