⚡ electrical engineering

Fault Diagnosis for Satellite of Multi-Subsystems with Digital Twin Residual

This paper proposes a Digital Twin Residual (DTR)-driven fault diagnostic method for multi-subsystem satellites that incorporates subsystem coupling and transforms raw data into residual features to overcome existing modeling limitations and achieve superior diagnostic performance.

Original authors: Yu Shi, Xuelei Deng, Yunfeng Dong, Fengrui Liu

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yu Shi, Xuelei Deng, Yunfeng Dong, Fengrui Liu

Original paper licensed under CC BY 4.0 (https://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

Imagine a satellite not as a cold, silent machine floating in space, but as a bustling, high-tech city with seven different neighborhoods: the Attitude and Orbit Control (the steering wheel), the Power Grid (solar wings and batteries), the Thermal Control (the air conditioning), and others. In this city, everything is connected. If the steering wheel gets stuck, the solar wings might point the wrong way, which changes how much power they generate, which then messes up the battery charge, and suddenly the whole city is overheating.

For a long time, trying to find a broken part in this city was like trying to hear a whisper in a hurricane. Engineers would look at one neighborhood at a time, or they would try to listen to the raw noise of the sensors. But the "noise" of normal operations—like the satellite turning to face the sun or the battery charging—was so loud that it drowned out the actual "screams" of a broken part.

The Problem with the Old Way
The paper argues that the old methods were missing the big picture. They treated the satellite's neighborhoods as if they were isolated islands. If a problem happened in the Power Grid, the old models didn't realize it might be because the Steering Wheel neighborhood was acting up. Also, they tried to feed the computer raw data straight from the sensors. The authors found that this raw data is a messy mix of "normal life" and "broken parts," making it incredibly hard for a computer to learn what a real fault looks like. They explicitly ruled out the idea that just using raw data or looking at single subsystems is enough to solve this.

The New Superpower: The Digital Twin and the "Residual"
To fix this, the researchers built a Digital Twin. Think of this as a perfect, virtual clone of the satellite living on a computer on Earth. This clone knows all the rules of the city: how the steering wheel affects the solar wings, how friction heats up the wheels, and how the sun's angle changes the power.

But a clone isn't perfect; it gets a little out of sync with the real satellite over time, just like a map gets outdated. So, the team used a special math trick (called an NLS method) to constantly update the clone's settings using real data from the actual satellite. This made the virtual twin a perfect mirror of the real thing.

Here is the magic trick: The Residual.
Instead of listening to the satellite's noisy voice, the researchers asked the Digital Twin to predict what the satellite should be doing. Then, they subtracted that prediction from what the satellite actually did.

If everything is working, the prediction and the reality match perfectly. The difference (the residual) is zero. It's like silence.
If something breaks, the prediction fails. The difference (the residual) suddenly spikes.

The paper found that this "difference" is where the truth hides. By stripping away the normal background noise, the fault becomes a bright, clear signal. It's like turning off the music at a party so you can finally hear the person shouting, "The cake is burning!"

The Proof: Ladybird-1 and 49 Scenarios
The team tested this idea using a real satellite called Ladybird-1. They didn't just guess; they simulated 49 different disaster scenarios (like a stuck wheel, a broken sensor, or a short circuit) to create a massive training dataset.

They trained a smart computer brain (a Temporal Convolutional Network, or TCN) to look at these "residual" signals. The results were impressive:

When they fed the computer the messy raw data, it got confused. The data looked like a tangled ball of yarn.
When they fed it the clean residual data, the computer saw clear patterns. It was like the tangled yarn suddenly untangled itself into neat, separate balls.

The new method, called DTR-TCN, achieved a 96.02% accuracy rate in spotting faults. Even better, it had 0% false alarms (it never cried wolf when there was no wolf) and 0% false negatives (it never missed a real wolf). In comparison, other methods they tested missed faults or sounded false alarms.

Why This Matters
The paper suggests that this approach is a game-changer for making space systems "resilient"—meaning they can keep working even when things go wrong. Because the computer can spot the fault instantly (in just 0.34 milliseconds per batch of data), it can react fast enough to save the satellite.

The authors are careful to note that this was tested on a specific dataset of 49 scenarios and a specific satellite model. While the results are strong, they suggest that future work needs to look at even stranger, unknown faults and different types of satellites. But for now, they have shown that by building a perfect virtual twin and listening to the "silence" between the real and the virtual, we can finally hear the satellite's secrets clearly.

Technical Summary: Fault Diagnosis for Multi-Subsystem Satellites with Digital Twin Residual

1. Problem Statement

Fault diagnosis for multi-subsystem satellites is essential for constructing resilient space architectures, yet current Digital Twin (DT)-driven diagnostic methods face two primary limitations:

Insufficient Modeling of Subsystem Coupling: Existing DT models often fail to adequately capture the complex physical quantity exchanges and mutual influences between different satellite subsystems (e.g., Attitude and Orbit Control, Power, and Thermal Control). This lack of comprehensive coupling modeling hinders the accurate simulation of fault propagation across the system.
Difficulty Handling Complex Data Features: Satellite fault data contains a mixture of operational characteristics (normal behaviors like attitude maneuvers) and fault characteristics (abnormal patterns). Current methods typically train Deep Neural Networks (DNNs) on raw fault data or expert-preprocessed data, which leads to performance bottlenecks as the networks struggle to isolate specific fault features from the complex operational background. Additionally, data scarcity remains a challenge, as physical fault experiments are prohibitively expensive.

2. Methodology

The paper proposes a Digital Twin Residual (DTR) driven fault diagnostic method, implemented through a five-step framework:

A. Multi-Subsystem Digital Twin Modeling

A unified satellite model is constructed to simulate the coupling relationships between seven primary subsystems: Attitude and Orbit Control (AOCS), Power, Thermal Control, Propulsion, Telemetry and Command (TTC), Structural, and Payload.

Coupling Mechanism: The model explicitly calculates physical quantity exchanges. For example, the torque of a reaction wheel (AOCS) is coupled to the power subsystem via input voltage and to the thermal subsystem via friction-induced temperature changes. Similarly, solar wing power output is coupled to orbital position and attitude angles.
Virtual-Physical Interaction: To ensure the DT model reflects the actual satellite's evolving state (due to aging or degradation), a Non-Linear Least Squares (NLS) method is employed. This iteratively updates the model's initial state variables using real telemetry data from the Ladybird-1 satellite, minimizing the Root-Mean-Square Error (RMSE) between simulation and telemetry.

B. Residual Data Construction

Instead of using raw fault state data, the method generates residual data.

Process: The DT model simulates both normal and fault states under various operational conditions. Residuals are calculated as the difference between the simulated normal state data and the simulated fault state data ( $y_{residual} = y_{fault} - y_{normal}$ ).
Rationale: This process decouples operational features (which are present in both states) from fault-related features. The resulting residual data isolates the fault signatures, making them more distinct and reducing feature complexity.

C. Fault Diagnosis via DTR-TCN

A Temporal Convolutional Network (TCN) is trained on the constructed residual dataset.

Architecture: The network includes a convolutional layer, stacked TCN blocks (utilizing dilated causal convolutions to capture long-term dependencies), an adaptive max pooling layer, and a fully connected layer.
Training: The network is trained to map residual samples to specific fault categories (49 scenarios in total, including 48 fault types and 1 normal state) by minimizing cross-entropy loss.

3. Key Contributions

Methodological Framework for Multi-Subsystem DT: The paper proposes a framework that integrates subsystem coupling relationships into a unified model. It clarifies the mechanism of coupling through physical quantity exchange and supports virtual-physical interaction via NLS-based model updating, enabling the simulation of coupled parameter variations under fault scenarios.
Digital Twin Residual (DTR) Diagnostic Method: The study introduces a method that transforms raw fault data into residual data to decouple operational features from fault features. This approach enhances the diagnostic capability of DNNs by providing input data where fault characteristics are significantly more prominent.
Validation on a Multi-Subsystem Dataset: The method is validated on a dataset containing 49 fault scenarios derived from the Ladybird-1 satellite, demonstrating superior performance over existing DT-driven and classical deep learning methods.

4. Experimental Results

The proposed DTR-TCN was evaluated against several existing methods (CNN, DCBA, DRSN, BLC-CNN, and LSTM) using a dataset of 4,900 samples (49 scenarios × 100 samples).

Feature Visualization: t-SNE visualization revealed that raw fault data exhibited a highly mixed distribution with low discriminability. In contrast, residual data showed clear clustering tendencies with improved separation between different fault scenarios.
Diagnostic Performance:
- Accuracy: The DTR-TCN achieved an accuracy rate of 96.02% and an F1-score of 95.91%.
- Error Rates: The method achieved 0.00% for both False Alarm Rate and False Negative Rate, outperforming all comparison methods which exhibited non-zero errors in one or both categories.
- Comparison: Compared to the DCBA method, the DTR-TCN improved accuracy by 2.95% and F1-score by 3.21%.
Computational Efficiency: The DTR-TCN demonstrated high efficiency with 61,041 parameters, an average training time of 1.78 seconds per epoch, and an inference time of 0.34 ms per batch. This meets real-time engineering requirements, as the inference time is significantly shorter than typical telemetry sampling intervals.

5. Significance and Claims

The paper claims that the proposed DTR method offers a robust solution for DT-driven satellite fault diagnosis by addressing the critical challenges of subsystem coupling and data feature complexity.

Resilience: By enabling accurate diagnosis across multiple coupled subsystems, the method contributes to the resilience of space architectures, ensuring continuous service even under fault conditions.
Data Efficiency: The approach leverages the DT model to generate cost-effective fault data, overcoming the scarcity of real-world fault measurements.
Reliability: The elimination of false alarms and false negatives in the experimental validation suggests high reliability for practical deployment, reducing the need for manual verification by operators during long-duration normal operations.

The authors conclude that this study provides a reference for future DT-driven satellite fault diagnosis, with future work planned to address unknown/compound faults and cross-platform diagnostic applications.