Michael Herman
(Advisor: Dr. Dimitri Mavris)
will defend a doctoral thesis entitled,
Predictive calibration for digital sun sensors using sparse submanifold convolutional neural networks
On
Thursday, December 12 at 9:00am EST
Collaborative Visualization Environment (CoVE)
Weber Space Science and Technology Building (SST II)
And on
Abstract
Recent developments of AI techniques for space applications mirror the success achieved in terrestrial applications. Machine learning, which scales in data rich environments, is particularly well suited to space-based computer vision applications such as space optical attitude sensing. Attitude sensors determine the spacecraft attitude through the sensing of an astronomical object, in which the Sun and fixed stars are the two primary astronomical sensing objects. Attitude sensors are critical components for the survival and knowledge improvement of spacecraft. Of these, digital sun sensors are the most common and important sensor for spacecraft attitude determination. Nearly all low-Earth orbiting small satellites employ sun sensors as part of the attitude sensor package, which determine the satellite attitude by measuring the Sun vector relative to the satellite coordinates.
The operation of small satellites requires highly accurate and reliable sensing techniques for attitude determination. Generally, performance critical missions are supported by star trackers, however star trackers are not always practical for small satellite operations. The main challenge in using sun sensors for attitude estimation is sensor errors, which limit the overall achievable estimation accuracy. However, the traditional sun sensor calibration process is costly, slow, labor intensive and inefficient. Furthermore, conventional sun sensor models are architecture specific, require subject matter expert (SME) knowledge to develop, need both a feature extraction and regression step, and are limited by centroiding accuracy. These limitations motivate the use of AI techniques to enable more accurate and efficient calibration.
Star trackers, Earth cameras, and sun sensors have all demonstrated improvements from the application of machine learning techniques. However, limited coverage has been provided to sun sensor calibration. While deep learning has been investigated to calibrate sun sensors, these methods have been focused on analog sensors. Unlike analog sensors, digital sun sensor features are inherently high-dimensional and sparse, thereby making training on dense networks infeasible. These challenges necessitate the use of sparse convolutional neural networks (SCNN) in this work.
The objective of this dissertation is to develop an end-to-end predictive calibration methodology for digital sun sensors to solve 2-axis state estimates utilizing a sparse submanifold convolutional neural network (SSCNN). A methodology is developed to address the gaps in traditional digital sun sensor calibration through: (i) generation and augmentation of datasets, (ii) modeling and training, and (iii) assessment of model credibility. In the first step, synthetic data is generated with a physics-informed simulation and then augmented. The second step involves the development and training of a novel predictive calibration model via convolutional neural network to address the limitations of conventional methods. The model trained is a modified ResNet-34 based regression SSCNN to solve for the two sun angles. Finally, the third step assesses the model credibility using verification and validation, model testing, and an evaluation of the model robustness.
The proposed method solves the following challenges of traditional calibration: (1) a SME is not required to formulate the model, (2) a separate feature extraction and model is not required, (3) no segmentation is required, (4) richer feature extraction is enabled, (5) the method is mask agnostic, (6) no denoising is required, and (7) sub-FOVs are accounted for via classes. We find that the proposed method achieves state-of-the-art performance with a mean accuracy of 0.005 deg for the two sun angle estimates. Furthermore, the model is highly capable of implicitly learning complex noise patterns and handling mixed noise types, thereby greatly improving the model robustness and accuracy to real-world applications. The main contributions of this work are: (1) the first time to our knowledge a CNN regression model is applied to the problem of sun sensor predictive calibration, (2) introducing a fused end-to-end training approach for digital sun sensor calibration, (3) creating a publicly available physics-informed synthetic dataset and simulation for digital sun sensor training images, and (4) conducting a performance evaluation of the deep learning approach to various mask configurations.
Committee
- Prof. Dimitri Mavris – School of Aerospace Engineering (advisor)
- Prof. Glenn Lightsey – School of Aerospace Engineering
- Prof. John Christian – School of Aerospace Engineering
- Prof. Duen Horng (Polo) Chau – School of Computational Science and Engineering
- Dr. Olivia Pinon Fischer – School of Aerospace Engineering