🤖 machine learning

Early Prediction of Satellite Collision Probability Using a Hybrid TCN-Transformer Model for a CDM-Based Conjunction Analysis Framework

This paper proposes a hybrid TCN-Transformer model that leverages sensitivity analysis and Principal Component Analysis on Conjunction Data Messages (CDMs) to accurately predict future collision probabilities for Low Earth Orbit satellites, thereby enabling earlier and more consistent risk assessment for conjunction avoidance maneuvers.

Original authors: Rabia Tüylek Tok, Burak Yağlıoğlu, Enes Dağ, Emre Onur Kahya

Published 2026-09-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rabia Tüylek Tok, Burak Yağlıoğlu, Enes Dağ, Emre Onur Kahya

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The sky above us is becoming increasingly crowded. While the night once held only the steady drift of natural stars, low Earth orbit is now filled with thousands of active satellites and countless fragments of debris left behind by decades of space exploration. These objects travel at tremendous speeds, so fast that even a tiny speck of paint could shatter a solar panel or disable a spacecraft upon impact. To keep these valuable machines safe, operators constantly monitor the paths of every object, looking for moments when two of them might pass dangerously close to one another. When such a close approach is predicted, a specific warning message is generated, containing a detailed snapshot of the situation: where the objects are, how fast they are moving, and a calculated estimate of the chance they will collide.

However, this warning system has a built-in delay. The messages are updated over time as new tracking data arrives, and the estimated chance of collision can swing wildly from one update to the next. This is because the uncertainty about where a satellite actually is grows as time passes, much like a fog thickening around a moving car. For satellites that rely on electric propulsion, which can only push gently and slowly, waiting for the final, most accurate warning message can be dangerous. By the time the last message arrives, there may not be enough time left to plan and execute a safe maneuver. The challenge, then, is to look at the early, imperfect warnings and predict how the risk will evolve before the next update arrives.

A team of researchers at TUBITAK UZAY and Istanbul Technical University has developed a new way to solve this problem. Instead of waiting for the final data, they created a system that learns to anticipate the future risk based on the sequence of early warnings. They treated the stream of warning messages as a story that unfolds over time, where each new message changes the plot. To teach a computer how to read this story, the researchers first had to understand exactly how the physics of space travel affects the numbers in those messages. They used a sophisticated method to simulate how tiny changes in a satellite's position or speed would ripple through the system, causing the uncertainty about its location to grow or shrink. They found that even a minuscule one percent shift in the input data could cause the calculated uncertainty to jump by as much as sixty-eight percent, proving that the system is incredibly sensitive to small details.

With this understanding of the physics, the team built a rich dataset that went beyond the raw numbers found in the standard warning messages. They added new, calculated values that described the shape of the uncertainty, the angle at which the satellites were approaching each other, and the direction of their movement. They also looked at how these values were trending over time, noting whether the risk was rising or falling with each new message. This enriched information was then fed into a hybrid computer model that combines two powerful types of artificial intelligence. One part of the model is excellent at spotting short-term changes between consecutive messages, while the other part is designed to understand the long-term patterns across an entire sequence of events. Together, they learn to recognize the subtle signatures that indicate whether a close approach will turn into a collision or simply pass safely.

When the researchers tested this system on real-world data from TUBITAK UZAY, the results were promising. The model was able to look at the first few warning messages for a specific event and accurately predict the collision probability that would appear in the very next message. It successfully handled a wide range of scenarios, from extremely dangerous situations where the chance of impact was high, to very safe passes where the risk was nearly zero. In many cases, the predicted risk was almost identical to the actual risk that appeared later. Crucially, the system learned to distinguish between events that truly needed attention and those that were harmless, correctly identifying dangerous situations about eighty-one percent of the time while rarely raising false alarms for safe events.

This approach offers a new layer of safety for satellite operators. By providing an early estimate of how the risk will change, the system gives operators a head start. For satellites with slow-moving electric engines, this extra time could be the difference between a successful avoidance maneuver and a catastrophic collision. The study suggests that by combining a deep understanding of orbital physics with modern machine learning, we can turn a stream of uncertain data into a reliable forecast, allowing us to navigate the crowded skies of the future with greater confidence and safety.

Technical Summary: Early Prediction of Satellite Collision Probability Using a Hybrid TCN–Transformer Model

Problem Statement
The rapid expansion of operational satellites and orbital debris in Low Earth Orbit (LEO) has intensified the frequency of close approach events, creating significant operational burdens for satellite operators. This challenge is particularly acute for satellites utilizing electric propulsion, where low-thrust capabilities necessitate longer maneuver execution times compared to chemical propulsion. Consequently, operators require earlier risk assessment to initiate planning before the final Conjunction Data Message (CDM) becomes available.

Current operational practice relies on CDMs, which provide relative state vectors, covariance matrices, miss distance, time of closest approach (TCA), and the estimated probability of collision (PoC). However, interpreting sequential CDMs is difficult because the reported PoC is highly sensitive to the time evolution of orbital uncertainty. As the satellite state is propagated, the covariance matrix expands due to nonlinear orbital dynamics and perturbations, altering the uncertainty ellipsoid and the resulting collision probability. Existing workflows typically evaluate the latest available CDM without explicitly estimating how the risk metric will evolve in the subsequent update, potentially delaying critical decision-making.

Methodology
The study proposes a learning-based framework that combines physically grounded sensitivity analysis with a hybrid deep learning architecture to predict the PoC of a future CDM update based on historical sequential data.

Sensitivity Analysis and Dataset Construction:
- Unscented Transform (UT) Propagation: To quantify the sensitivity of the collision risk metric to CDM parameters, the authors employ an Unscented Transform-based propagation and backpropagation framework. A controlled 1% variation is applied individually to state vector components (position and velocity) and physical parameters (drag coefficient $C_d$ , solar radiation pressure coefficient $C_r$ , and cross-sectional area $A$ ). The state is backpropagated 24 hours and then forward-propagated to the original epoch using an RK8 integrator within a nonlinear dynamical model (including $J_2$ , drag, SRP, and third-body effects).
- Sensitivity Metrics: The resulting covariance matrices are compared to the baseline using the Frobenius norm. Results indicate that a 1% input variation leads to covariance changes ranging from approximately 30% to 68%, confirming high sensitivity.
- Principal Component Analysis (PCA): PCA is applied to raw CDM parameters to identify features most relevant to PoC variation.
- Enriched Dataset Generation: The final dataset consists of "enriched conjunction data" samples. Each sample combines 31 raw CDM parameters (e.g., relative position/velocity in RTN frame, covariance terms, physical coefficients) with 13 derived physical metrics. These derived metrics include the 2D collision plane distance, 3D covariance ellipsoid volume, Mahalanobis distance, approach angle, out-of-plane ratio, and trend-based features (slopes of PoC, miss distance, and relative motion over time).
Hybrid TCN–Transformer Model:
- Architecture: The model is designed to handle the sequential nature of CDMs (organized by Event ID). It utilizes a Temporal Convolutional Network (TCN) to capture short-term temporal variations and local transitions between consecutive CDMs. A Transformer encoder follows the TCN to capture long-range dependencies across the entire sequence, learning which historical updates are most critical for predicting the next step.
- Input/Output: The input is a chronological sequence of enriched conjunction data vectors ( $k$ samples). The output is the predicted PoC value for the subsequent ( $k+1$ ) CDM update.
- Training: The model is trained using a Smooth L1 loss function to minimize the difference between true and predicted PoC values. The dataset is partitioned by Event ID (79.63% training, 8.00% validation, 12.37% test) to ensure realistic evaluation.

Key Results
The framework was applied to real CDM data received and analyzed within T ÜB İTAK UZAY.

Prediction Accuracy: The hybrid model successfully reproduced the general magnitude and temporal trends of future PoC values across a wide numerical range, from high-risk cases ( $10^{-3}$ – $10^{-4}$ ) to very low-risk cases ( $10^{-7}$ – $10^{-8}$ ). For several events, predicted values closely matched true values. In cases where the model slightly underestimated high-risk values, the predictions remained within the same risk regime and order of magnitude.
Zero-Risk Handling: The model correctly predicted zero PoC for events with no estimated collision probability, avoiding the introduction of artificial risk.
Classification Performance: When evaluated as a binary classification task (Risk vs. No-Risk, threshold $10^{-6}$ $1 0^{- 6}$ ), the model achieved:
- Accuracy: 0.9398
- Recall: 0.8125 (indicating strong detection of risky events)
- Specificity: 0.9572 (indicating strong identification of non-risk events)
- F1-score: 0.7647
- F2-score: 0.7926 (suggesting the model prioritizes recall, which is critical for safety).

Significance and Claims
The paper claims that this framework provides a practical method for earlier and more consistent operational risk evaluation for LEO satellite conjunctions. By estimating the PoC of the next CDM update before it is available, the system supports operational planning for satellites with limited maneuver capabilities (such as those using electric propulsion).

The study emphasizes that the framework does not replace the final CDM assessment but serves as an additional decision-support layer. It enables operators to identify events requiring closer monitoring or preliminary maneuver preparation earlier in the conjunction timeline. The integration of physically derived metrics (via UT and PCA) with deep learning allows the model to learn the complex, nonlinear evolution of orbital uncertainty and collision risk from sequential data.

Future Work
The authors note that future work will extend the framework using ensemble learning to improve robustness and generalization. Additionally, they plan to investigate the prediction of other CDM-related parameters, such as relative motion and encounter geometry metrics, to create a more comprehensive early conjunction risk assessment framework.

More like this