A One-Class Support Vector Machine Calibration Method for Time Series Change Point Detection

Baihong Jin, Yuxin Chen, Dan Li, Kameshwar Poolla, Alberto L. Sangiovanni-Vincentelli

Abstract: It is important to identify the change point of a system's health status, which usually signifies an incipient fault under development. The One-Class Support Vector Machine (OC-SVM) is a popular machine learning model for anomaly detection and hence could be used for identifying change points; however, it is sometimes difficult to obtain a good OC-SVM model that can be used on sensor measurement time series to identify the change points in system health status. In this paper, we propose a novel approach for calibrating OC-SVM models. The approach uses a heuristic search method to find a good set of input data and hyperparameters that yield a well-performing model. Our results on the C-MAPSS dataset demonstrate that OC-SVM can also achieve satisfactory accuracy in detecting change point in time series with fewer training data, compared to state-of-the-art deep learning approaches. In our case study, the OC-SVM calibrated by the proposed model is shown to be useful especially in scenarios with limited amount of training data.

The Challenge of Change Point Detection

In the operation of Cyber-Physical Systems (CPS), such as jet engines, identifying the exact moment a system shifts from a healthy state to an incipient fault state is critical. This transition point, often referred to as the "knee" of the degradation curve, represents the optimal window for predictive maintenance. Acting before this point results in wasted resources, while acting too late leads to catastrophic failure.

The "knee" of the degradation curve, representing the transition from normal operation to accelerated failure.

Detecting this "knee" is difficult because precise labels for when a fault begins are rarely available in real-world sensor data. While unsupervised tools like One-Class Support Vector Machines (OC-SVM) are effective at spotting anomalies, they are highly sensitive to hyperparameters, such as the kernel width $\gamma$. Traditionally, tuning these settings requires labeled data, creating a paradox for unsupervised applications.

Methodology: Self-Consistent Calibration

To address the lack of labels, this paper introduces a "self-consistent" calibration framework. Instead of manually selecting model settings, the approach treats the location of the change point in the training data as a variable to be optimized alongside model hyperparameters.

The Model and Optimization

The framework utilizes an OC-SVM with a Radial Basis Function (RBF) kernel. To find the optimal configuration, the authors employ Differential Evolution (DE), a gradient-free heuristic search algorithm. DE simultaneously searches for:

  1. The hyperparameter $\gamma$.
  2. A vector $\rho$, representing the hypothesized "healthy" portion of the life cycle for each training engine.

The Loss Function

The core innovation is the loss function, which measures the "self-consistency" of the model. If a model trained on an assumed 70% healthy data split subsequently predicts that exactly those first 70% of cycles are healthy, the hypothesis is considered self-consistent and likely reflects the true physical degradation state.

Experimental Results

The method was validated using the C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) dataset, specifically the challenging FD004 subset.

Data Efficiency

A key finding of this research is the extreme data efficiency of the calibrated OC-SVM. The proposed method achieved high accuracy using only 20 engine instances for training. In comparison, state-of-the-art deep learning models like the WTTE-RNN require significantly larger datasets (e.g., 249 engines) to produce stable change point detections.

Detected change points (blue dots) for the C-MAPSS FD004 dataset, showing alignment with the transition region.

Validation via Uncertainty

As shown in the figure above, the OC-SVM's detected change points (blue dots) align closely with the transition regions identified by more complex models. By optimizing both the data split and hyperparameters (finding an optimal $\gamma \approx 3.59$), the OC-SVM effectively identifies the start of accelerated degradation without requiring pre-existing fault labels.

Conclusion

This research demonstrates that well-calibrated "classical" machine learning models can outperform data-hungry deep learning architectures in scenarios where training data is scarce. By treating the change point as an optimizable parameter, the proposed heuristic calibration method provides a robust, label-free solution for predictive maintenance in complex industrial systems. Future work will explore alternative heuristic search methods, such as Bayesian Optimization, to further improve calibration efficiency.

Summary updated

Abstract

It is important to identify the change point of a system's health status, which usually signifies an incipient fault under development. The One-Class Support Vector Machine (OC-SVM) is a popular machine learning model for anomaly detection and hence could be used for identifying change points; however, it is sometimes difficult to obtain a good OC-SVM model that can be used on sensor measurement time series to identify the change points in system health status. In this paper, we propose a novel approach for calibrating OC-SVM models. The approach uses a heuristic search method to find a good set of input data and hyperparameters that yield a well-performing model. Our results on the C-MAPSS dataset demonstrate that OC-SVM can also achieve satisfactory accuracy in detecting change point in time series with fewer training data, compared to state-of-the-art deep learning approaches. In our case study, the OC-SVM calibrated by the proposed model is shown to be useful especially in scenarios with limited amount of training data.
BibTeX
@inproceedings{jin2019one,
  author = {Baihong Jin and Yuxin Chen and Dan Li and Kameshwar Poolla and Alberto L. Sangiovanni-Vincentelli},
  title = {{A One-Class Support Vector Machine Calibration Method for Time Series Change Point Detection}},
  booktitle = {ICPHM},
  year = {2019},
  doi = {10.1109/icphm.2019.8819385},
}
Figure