Back to Projects
Reliability Engineering

Predictive Maintenance

View Source Code

Project Overview

Forecasting component degradation trends across satellite fleets to optimize maintenance schedules, de-orbiting plans, and operational loads.

The Problem

Operators lacked the tools to dynamically predict when a component would fail based on its actual usage and environmental stress. This led to either premature decommissioning of healthy satellites or unexpected catastrophic failures in orbit.

Our Solution

We implemented a Bayesian deep learning framework that not only predicts future states of degrading components (like batteries and solar arrays) but also provides a mathematically rigorous uncertainty bound for every prediction.

System Architecture

The pipeline utilizes recurrent neural networks (LSTMs) coupled with Bayesian inference layers. Data is streamed continuously from orbit into a cloud-native Snowflake data warehouse, where scheduled Airflow dags run the predictive models.

Impact & Results

Increased the average operational lifespan of monitored satellites by 18 months. Saved $45M in deferred launch costs by safely extending the life of three aging communications satellites.

AI Models

Bayesian LSTMsMarkov Chain Monte Carlo (MCMC)Gaussian Processes

Technology Stack

SnowflakeApache AirflowPyMC3PythonAWS SageMaker