Method and device for training and predicting a conjunction parameter from conjunction data messages
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting position uncertainty at the time of closest approach in space object collisions lack accuracy and interpretability, particularly for secondary objects like debris, leading to sub-optimal collision avoidance strategies and increased operational costs.
Innovation Solution
A Transformer-based machine-learning model is trained using conjunction data messages (CDMs) to forecast position uncertainty, employing self-attention mechanisms and positional encoding to capture long-range dependencies in irregularly spaced time series data, enabling robust and interpretable predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning methods (decision trees, neural networks) are used to predict position uncertainty, then the prediction task can be performed, but the accuracy and interpretability are insufficient
Solution Approach 1:
The patent replaces traditional machine learning approaches (decision trees, basic neural networks) with a Transformer-based model that uses self-attention mechanisms. This substitution enables the model to capture long-range dependencies in conjunction data messages while providing interpretable attention weights that show which input features most influence predictions, thereby improving both accuracy and interpretability simultaneously
Solution Approach 2:
The patent transforms the input conjunction data messages by applying positional encodings and embedding layers, changing the parameter representation from raw numerical values to enriched feature vectors. This parameter transformation allows the Transformer model to effectively process irregularly spaced time series data while maintaining interpretability through the attention mechanism
2Measurement precision
If predictions are made close to the critical decision moment for accurate time of closest approach, then prediction accuracy improves, but safe avoidance maneuvers may not be feasible or incur significant costs
Solution Approach 1:
The patent enables operators to perform preliminary assessments of position uncertainty evolution well before the critical decision moment. By forecasting uncertainty at multiple future time points using the Transformer model, operators can evaluate potential collision risks in advance and plan avoidance maneuvers proactively, rather than reacting at the last moment when accurate predictions are available but maneuver options are limited
Solution Approach 2:
The patent implements a feedback mechanism where the Transformer model continuously forecasts position uncertainty evolution based on incoming conjunction data messages. This feedback loop allows operators to monitor predicted uncertainty trends over time and adjust their risk assessment and maneuver planning accordingly, improving decision-making both in terms of timing and accuracy
3Productivity
If the number of space objects increases, then more conjunctions need to be monitored, but the complexity of collision risk assessment increases
Solution Approach 1:
The patent develops a universal Transformer-based model that can process conjunction data messages for any pair of space objects regardless of their specific characteristics. This single model architecture handles diverse input scenarios (different object types, orbital regimes, data quality levels) uniformly, enabling scalable monitoring of increasing numbers of conjunctions without proportionally increasing assessment complexity
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present document discloses a computer-implemented method for predicting a conjunction parameter from conjunction data messages and the respective method for training a transformer-based machine-learning model for predicting a conjunction parameter. It is also disclosed a device comprising a computer-readable medium comprising the trained transformer-based machine-learning model and a computer configured to carry out the training method of the transformer-based machine-learning model.