Methodology for Correlating Historical Device Degradation Data to Radiation-Induced Degradation System Effects

Abstract

An engineering team needs a general understanding of their system’s failure probability, especially when COTS parts exist within the system. Deriving such a probability can be difficult if the spacecraft team lacks an expert who can provide a foundational understanding of potential radiation degradation being introduced into their system. Multiple rounds of radiation testing can also be prohibitive due to test facilities’ costs (potentially costing upwards of $1,000 per hour), and the time and resources required to prepare for testing. Therefore, a method is proposed for deriving a preliminary system-level failure probability from device-level radiation degradation datasets. This method converts device-level degradation probabilities (derived from historical datasets of similar devices) into a system-level failure probability through a Monte Carlo method. This historical data (which can be smaller/older/less relevant than ideal) can be found in radiation databases (e.g., PMPedia database, NASA Goddard Space Flight Center radiation database). While not a substitute for traditional radiation testing (e.g., MIL-STD-883 1019 for total ionizing dose (TID) effect testing), the presented methodology can provide a degree of confidence in a system’s reliability. These probabilities can then be incorporated into a system’s fault tree to reveal devices most at-risk of degradation. This process can also be used for analyzing system-led errors and predicting probability of system failure. As a demonstration system for this methodology, a simple LiDAR system is utilized because it can be simply modelled using a circuit simulator (e.g., LTspice [9]) and a fault tree generator (e.g., Command-line Risk Analysis Multi-tool (SCRAM) [10]). The methodology can be summarized as follows: device-level data is feed into both kernel density estimation (KDE) and a Bayesian statistical model. The KDE process results in the most likely distribution shape of the dataset, which is then inserted into the Bayesian model so that device-level failure probability distributions can be derived and inputted into a LTspice model of the system. The LTspice-simulated output of the modeled LiDAR system is then run 50 times, and the system-level outputs are converted into a system-level failure probability.