Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Image-based Detection of Segment Misalignment in Multi-mirror Satellites using Transfer Learning

C. Tanner Fredieu, Jonathan Tesch, Andrew Kee, David Redding

TL;DR

Problem: segment misalignment in multi-mirror satellites poses maintenance challenges in space. Approach: a transfer-learning framework using pre-trained image models on FFT-transformed grayscale patches detects ghost components and estimates misalignment intensity. Key findings: binary misalignment detection accuracy of up to $98.75\%$ and intensity estimation accuracy of $98.05\%$, achieved with a patch-based FFT preprocessing on simulated 4-, 6-, and 8-mirror CubeSats; the method enables a per-image confidence metric. Significance: the results suggest a software-defined, autonomous health-monitoring capability for next‑generation, segmented-mirror spacecraft.

Abstract

In this paper, we introduce a system based on transfer learning for detecting segment misalignment in multimirror satellites, such as future CubeSat designs and the James Webb Space Telescope (JWST), using image-based methods. When a mirror segment becomes misaligned due to various environmental factors, such as space debris, the images can become distorted with a shifted copy of itself called a "ghost image". To detect whether segments are misaligned, we use pre-trained, large-scale image models trained on the Fast Fourier Transform (FFT) of patches of satellite images in grayscale. Multi-mirror designs can use any arbitrary number of mirrors. For our purposes, the tests were performed on simulated CubeSats with 4, 6, and 8 segments. For system design, we took this into account when we want to know when a satellite has a misaligned segment and how many segments are misaligned. The intensity of the ghost image is directly proportional to the number of segments misaligned. Models trained for intensity classification attempted to classify N-1 segments. Across eight classes, binary models were able to achieve a classification accuracy of 98.75%, and models for intensity classification were able to achieve an accuracy of 98.05%.

Image-based Detection of Segment Misalignment in Multi-mirror Satellites using Transfer Learning

TL;DR

Problem: segment misalignment in multi-mirror satellites poses maintenance challenges in space. Approach: a transfer-learning framework using pre-trained image models on FFT-transformed grayscale patches detects ghost components and estimates misalignment intensity. Key findings: binary misalignment detection accuracy of up to and intensity estimation accuracy of , achieved with a patch-based FFT preprocessing on simulated 4-, 6-, and 8-mirror CubeSats; the method enables a per-image confidence metric. Significance: the results suggest a software-defined, autonomous health-monitoring capability for next‑generation, segmented-mirror spacecraft.

Abstract

In this paper, we introduce a system based on transfer learning for detecting segment misalignment in multimirror satellites, such as future CubeSat designs and the James Webb Space Telescope (JWST), using image-based methods. When a mirror segment becomes misaligned due to various environmental factors, such as space debris, the images can become distorted with a shifted copy of itself called a "ghost image". To detect whether segments are misaligned, we use pre-trained, large-scale image models trained on the Fast Fourier Transform (FFT) of patches of satellite images in grayscale. Multi-mirror designs can use any arbitrary number of mirrors. For our purposes, the tests were performed on simulated CubeSats with 4, 6, and 8 segments. For system design, we took this into account when we want to know when a satellite has a misaligned segment and how many segments are misaligned. The intensity of the ghost image is directly proportional to the number of segments misaligned. Models trained for intensity classification attempted to classify N-1 segments. Across eight classes, binary models were able to achieve a classification accuracy of 98.75%, and models for intensity classification were able to achieve an accuracy of 98.05%.
Paper Structure (13 sections, 6 equations, 6 figures, 4 tables)

This paper contains 13 sections, 6 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Non-machine Learning Methods
  4. Laplacian Kernel Method
  5. Fast Fourier Transform Method
  6. Machine Learning Methods
  7. Implementation
  8. Experimental Setup
  9. Datasets
  10. Model Architecture and System Design
  11. Experimental Results
  12. Future Work
  13. Conclusions

Figures (6)

  • Figure 1: Satellite image before and after FFT
  • Figure 2: Simulated arbitrary mirror configurations for 4, 6, and 8 mirrors
  • Figure 3: Image before and after ghosting
  • Figure 4: Final system design
  • Figure 5: Patch image being taken from satellite image
  • ...and 1 more figures