A probabilistic method for certification of analytically redundant systems

Bin Hu, Peter J. Seiler

Abstract

This paper introduces a rigorous mathematical framework to certify the reliability of analytically redundant aerospace systems—those that utilize software algorithms instead of additional hardware to detect faults. By applying the law of total probability over discrete time windows, the authors provide a method to calculate exact system failure rates. This approach demonstrates that sophisticated software can match or exceed the safety performance of heavy physical backups, providing a path toward meeting stringent aviation certification standards.

Introduction

In commercial aviation, critical systems must adhere to a strictly defined metric: a failure rate of no more than $10^{-9}$ per flight hour. Traditionally, engineers have met this requirement through physical redundancy, such as installing two or three backup actuators. However, physical redundancy is heavy, power-intensive, and expensive—posing a significant challenge for small Unmanned Aerial Vehicles (UAVs) and next-generation aircraft.

Analytical redundancy offers a solution by using mathematical models, known as Fault Detection and Isolation (FDI) logic, to monitor a single backup. The primary hurdle for this technology has been certification; aviation authorities like the FAA require rigorous proof of reliability. This paper bridges the gap by providing a standard method to translate FDI algorithm performance into the "failure per hour" metrics necessary for certification.

System Architecture

The authors focus on a duplex actuator system consisting of a primary actuator, a backup actuator, and an FDI logic component that determines when to switch between them. The architecture is designed to maintain system operation even if the primary component fails, provided the failure is detected and the backup is functional.

Block diagram of the duplex actuator system with analytical redundancy.

Methodology

Key to this framework is the analysis of reliability over an $N$-step time window rather than a single snapshot. For example, a one-hour flight segment at a 100Hz sampling rate involves 360,000 discrete steps. The authors decompose the probability of system failure into four mutually exclusive events using the Law of Total Probability:

  1. Missed Detection ($M_N$): The primary actuator fails, but the software does not detect it in time.
  2. False Alarm Induced Failure ($F_N$): The software incorrectly triggers a switch to the backup, which then fails.
  3. Proper Detection Dual Failure ($D_N$): The software correctly switches to the backup, but the backup fails anyway.
  4. Early False Alarm ($E_N$): The software switches to the backup before the primary fails, and the backup subsequently fails.

This decomposition allows for a simplified formula that separates hardware failure rates from FDI performance metrics, enabling engineers to identify the specific contributions of hardware and software to the overall safety budget.

Key Insights and Results

The research highlights that standard performance metrics like confusion matrices are insufficient for safety certification because reliability depends on how probabilities compound over thousands of time steps.

Performance Comparison

Numerical simulations based on standard aerospace parameters (MTBF of 1,000 hours, 100 Hz sample rate) showed that a well-tuned analytical redundancy system can actually outperform a triplex (three-actuator) physical system. In a scenario with a specific fault level ($f/\sigma = 10$), the dual-redundant system with analytical FDI achieved a failure probability of approximately $10^{-6}$ per hour, whereas a standard triplex hardware system had a higher failure probability of $3 \times 10^{-6}$.

Threshold Balancing

System reliability is extremely sensitive to the detection threshold ($H$). The paper demonstrates how to find an optimal "Goldilocks" zone—balancing the risks of false alarms and missed detections—using modified Receiver Operating Characteristic (ROC) curves.

Limitations and Future Work

While the proposed method provides a direct "plug-and-play" path for certification, several limitations remain:

  • Model Fidelity: The analysis assumes the FDI mathematical model is a perfect representation of reality. In practice, model uncertainty can degrade performance.
  • Independent Failures: The framework currently assumes actuators fail independently. Future work is needed to address "common-cause" failures (e.g., external impacts affecting both components).
  • Intermittent Faults: The model assumes that once a component fails, it remains failed. Real-world systems must handle intermittent glitches without permanently disabling healthy components.

The authors note that validating these probabilistic estimates against actual flight test data from small UAVs is the next critical step.

Summary updated

Abstract

Analytical fault detection algorithms have the potential to reduce the size, power and weight of safety-critical aerospace systems. Analytical redundancy has been successfully applied in many non-safety critical applications. However, acceptance for aerospace applications will require new methods to rigorously certify the impact of such algorithms on the overall system reliability. This paper presents a theoretical method to assess the probabilistic performance for an analytically redundant system. Specifically, a fault tolerant actuation system is considered. The system consists of dual-redundant actuators and an analytical fault detection algorithm to switch between the hardware components. The exact system failure rate per hour is computed using the law of total probability. This analysis requires knowledge of the failure rates for the hardware components. In addition, knowledge of specific probabilistic performance metrics for the fault detection logic is needed. Numerical examples are provided to demonstrate the proposed analysis method.
BibTeX
@article{hu2015probabilistic,
  author = {Bin Hu and Peter J. Seiler},
  title = {{A probabilistic method for certification of analytically redundant systems}},
  journal = {Int. J. Appl. Math. Comput. Sci.},
  year = {2015},
  doi = {10.1515/amcs-2015-0008},
}
Figure