Date of Award

8-6-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Daniel Berleant

Abstract

Predicting advancement in space exploration technology allows space agencies and companies to strategically allocate resources, prioritize missions, and explore collaborations on complex projects. This foresight acts as a roadmap for future exploration, making the best use of valuable resources. The present study employs spacecraft lifetimes as an exemplar of technological progress and proposes a novel forecasting model based on Long Short-Term Memory (LSTM) networks. Existing research often suggests an increasing exponential relationship between technology performance and time (the generalized Moore's Law). However, applying this model directly to spacecraft lifetimes has a limitation. Spacecraft lifespans are unknown at launch, and recent data is likely limited to short-lived missions. This creates a bias towards shorter lifespans for newer craft, especially with limited data. Excluding recent failures would not be the best solution as that data is valuable. This study addresses this limitation by introducing an end-time based function that utilizes recent failure data to provide more accurate forecasts. Furthermore, a new dataset is developed incorporating potential predictors of spacecraft lifetimes, such as launch weight and target destination. LSTM models are then trained on this dataset and fine-tuned with Bayesian optimization for the optimal hyperparameters to capture the complexities affecting spacecraft lifespans beyond the simple passage of time. Out-of-sample evaluation reveals the superiority of the optimized LSTM model compared to a baseline model. As our model goes beyond just time-based forecasting, this improvement highlights the significance of incorporating relevant features. This study represents a step towards improved forecasting of advancement in space exploration technology, as well as a methodology that could be applied to many other time series forecasting problems.

Download

Share

COinS