Performance Bounds for Finite Moving Average Tests in Transient Change Detection

Daniel Egea-Roca, José A. Lopez-Salcedo, Gonzalo Seco-Granados, H. Vincent Poor

Abstract

This paper deals with the problem of sequentially detecting statistical changes. In particular, the focus is on transient change detection, in which a probability minimizing optimal criterion is desirable. This is in contrast with the traditional minimization of the detection delay, proposed in quickest change detection problems. A finite moving average stopping time is proposed for the general transient change detection problem. The statistical performance of this stopping time is investigated and compared to other methods available in the literature. The proposed stopping time and theoretical findings are applied to quality monitoring, including reliability monitoring in industrial processes and signal quality monitoring in global navigation satellite systems. Numerical simulations are presented to assess the goodness of the presented theoretical results, and the performance of the considered stopping times. This will show the superiority of the proposed scheme.

Introduction: The Challenge of Transient Change Detection

In safety-critical applications like autonomous driving or urban navigation, GNSS "integrity"—the system's ability to provide timely warnings when position solutions are untrustworthy—is vital. While integrity standards are well-established for civil aviation, terrestrial environments face local threats like multipath interference, signal bouncing, and spoofing.

The mathematical challenge lies in the nature of these threats. Traditional Quickest Change Detection (QCD) assumes an anomaly persists indefinitely and aims to minimize detection delay. However, terrestrial threats are often transient (finite duration). In these cases, the goal is to ensure detection within a strict "Time to Alert" (TTA) window. If detection occurs after the TTA, it is considered a missed detection. Standard algorithms like CUSUM or the Shewhart test often fall short in this Transient Change Detection (TCD) framework because they are not optimized for finite-duration windows.

Methodology: The Finite Moving Average (FMA) Stopping Time

The authors propose a Finite Moving Average (FMA) stopping time approach. Unlike the CUSUM algorithm, which accumulates information from all past samples, the FMA only considers a sliding window of the last $m$ samples, where $m$ corresponds to the maximum allowable detection delay (the TTA).

Key Innovations

  • General Performance Bounds: The paper establishes tight upper bounds for the probability of missed detection ($P_{md}$) and false alarm ($P_{fa}$) for FMA. These bounds are valid for any statistical distribution, extending beyond simple Gaussian mean changes.
  • Signal-RAIM Integration: The FMA framework is applied to three GNSS metrics used for "Signal-Receiver Autonomous Integrity Monitoring" (sig-RAIM):
  • C/N0 (Carrier-to-Noise ratio): Modeled as a change in Gaussian mean.
  • DLLout (Code Discriminator Output): Modeled as a change in Gaussian variance.
  • SAM (Slope Asymmetry Metric): A complex case modeled as a simultaneous change in mean and variance.
ROC curves for Gaussian mean change (C/N0 metric), showing FMA's superior performance and the tightness of the derived theoretical bounds.

Results and Performance Analysis

Extensive Monte-Carlo simulations ($10^6$ runs) were used to validate the theoretical bounds. The simulations utilized parameters derived from established GNSS signal models to represent both "normal" and "multipath-affected" conditions.

Key Findings

  • Superior ROC Curves: In all tested metrics (C/N0, DLLout, and SAM), the FMA stopping time consistently outperformed CUSUM, Windowed CUSUM (WLC), and the Shewhart test. It achieved a lower probability of missed detection for any given false alarm rate.
  • Theoretical Validation: Simulated results closely matched the newly derived theoretical bounds, confirming that FMA performance is predictable and reliable for system design.
  • Practical Edge: In sig-RAIM examples, the FMA approach met strict integrity risk requirements (e.g., $10^{-2}$) where traditional methods like CUSUM failed, thereby increasing the "availability" of safety-critical services in challenging environments.
ROC curves for the complex Slope Asymmetry Metric (SAM), demonstrating FMA's robustness in non-standard statistical scenarios.

Conclusion and Future Directions

The "aha moment" of this research is that while CUSUM is optimal for infinite changes, the FMA is more natural for TCD because its memory matches the TTA requirements. Furthermore, the new FMA bounds are significantly tighter than previous bounds for windowed methods, allowing for more sensitive systems without increasing false alarms.

Future work includes providing a formal proof of FMA's optimality for the general non-Bayesian TCD problem across all distributions. Additionally, testing the FMA stopping time against real-world GNSS data from high-multipath urban environments is a logical next step for deployment.

Summary updated

BibTeX
@article{egearoca2018performance,
  author = {Daniel Egea-Roca and José A. Lopez-Salcedo and Gonzalo Seco-Granados and H. Vincent Poor},
  title = {{Performance Bounds for Finite Moving Average Tests in Transient Change Detection}},
  journal = {IEEE Trans. Signal Process.},
  year = {2018},
  doi = {10.1109/tsp.2017.2788416},
}
Figure