SAX Breakpoints for Random Forest Based Real-Time Contrast Control Chart

In-Seok Lee, Jun-Geol BaekAAAI 2019 · DOI

Introduction

Real-time fault detection is a cornerstone of modern manufacturing. Real-Time Contrast (RTC) control charts have emerged as a powerful tool, utilizing classifiers to distinguish between reference "normal" data and a moving window of current observations. While high-accuracy classifiers like Support Vector Machines are available, engineers often favor Random Forests (RF) for their inherent interpretability through variable importance metrics. However, RFs typically struggle with detecting subtle process shifts due to their discrete categorical outputs, which lack a continuous distance metric to the failure boundary. This paper proposes a novel data representation strategy using Symbolic Aggregate Approximation (SAX) with optimized breakpoints to overcome this limitation.

Methodology

The proposed approach integrates SAX to transform continuous time-series data into discrete symbols before classification. Standard SAX typically involves two steps: Piecewise Aggregate Approximation (PAA) to reduce dimensionality, followed by discretization based on Gaussian distribution breakpoints. The authors identify that standard Gaussian breakpoints are suboptimal for fault detection as they prioritize equal-area regions rather than the distribution tails where faults manifest.

The innovation lies in customizing these breakpoints using standard Z-scores. By adjusting the "topmost" breakpoint, the sensitivity of the RF classifier to outliers is significantly enhanced. The study evaluates three Z-score thresholds—1.6, 1.96, and 2.58—to identify the optimal configuration for process monitoring.

Experimental Setup

The methodology was validated using synthetic multivariate normal datasets with zero mean and identity covariance. To assess scalability and robustness, the experiments were conducted across two dimensions ($d=10$ and $d=100$) and two shift magnitudes ($\lambda = \sqrt{5}$ and $\sqrt{10}$). The primary performance metric was the Average Run Length ($ARL_1$), representing the number of observations required to detect a shift; lower values indicate superior performance.

Results and Discussion

The experimental results demonstrate that the choice of breakpoints is critical. As shown in the table below, the standard Gaussian-based SAX performed poorly, with $ARL_1$ values exceeding 147, compared to the original RTC's ~7.5. This confirms that "off-the-shelf" discretization is unsuitable for high-stakes fault detection.

Comparison of Average Run Length ($ARL_1$) across different methods and data dimensions.

In contrast, the proposed SAX 1.6 method consistently outperformed the original RTC chart. At $d=10$, SAX 1.6 achieved an $ARL_1$ of 7.40 (vs. 7.53 for RTC), and at $d=100$, it achieved 7.70 (vs. 7.93 for RTC). These results suggest that focusing on the distribution tails (the "1.6" threshold) makes the system more sensitive to process shifts without sacrificing the diagnostic benefits of the Random Forest model.

Conclusion

This research highlights that data representation can be as impactful as the choice of classification algorithm. By optimizing SAX breakpoints for the tails of the data distribution, the authors improved the detection speed of RF-based control charts. While the numerical gains over the original RTC appear modest, even small improvements in detection speed can lead to significant cost savings in high-speed manufacturing environments. Future work will focus on developing adaptive breakpoint strategies to handle non-Gaussian and non-linear real-world manufacturing data.

Summary updated

BibTeX
@inproceedings{lee2019sax,
  author = {In-Seok Lee and Jun-Geol Baek},
  title = {{SAX Breakpoints for Random Forest Based Real-Time Contrast Control Chart}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019959},
}
Figure