Empowering All-in-Loop Health Management of Spacecraft Power System in the Mega-Constellation Era via Human-AI Collaboration

This paper addresses the challenges of health management for spacecraft power systems in the emerging mega-constellation era by proposing the "Aligning Underlying Capabilities" principle and introducing SpaceHMchat, an open-source Human-AI collaboration framework validated on a realistic hardware platform and a new large-scale dataset to achieve high-precision, interpretable, and efficient all-in-loop health management.

The Gist

Imagine the future of space travel isn't just about sending one or two rockets to Mars, but launching thousands of tiny satellites to form a giant, global internet blanket (like Starlink or China's GW constellation). This is the "Mega-Constellation Era."

The problem? Managing thousands of satellites is a nightmare.

Right now, if a satellite's power system acts up, a team of highly paid, super-smart human experts has to wake up, read thousands of pages of manuals, check complex data charts, and figure out what's wrong. If you have 40,000 satellites, you'd need 40,000 teams of experts. That's impossible. It's too expensive, too slow, and there aren't enough humans to do it.

The Solution: SpaceHMchat (The "Space Co-Pilot")

This paper introduces a new system called SpaceHMchat. Think of it not as a robot that replaces humans, but as a super-smart AI assistant that works alongside humans. It's a "Human-AI Collaboration" tool designed to handle the boring, complex, and data-heavy parts of satellite health management, so the human experts can focus on the big, tough decisions.

Here is how it works, broken down into four simple steps using everyday analogies:

1. The "Traffic Cop" (Work Condition Recognition)

• The Old Way: A human looks at a dashboard with 50 different lights and gauges, trying to remember a complex rulebook to decide if the satellite is "charging," "draining," or "sleeping."
• The SpaceHMchat Way: You just ask the AI, "What is the satellite doing right now?" The AI acts like a Traffic Cop. It looks at the data, follows a strict set of logical rules (like a flowchart), and instantly says, "It's sunny, the batteries are charging, and no tasks are running." It does this in seconds with 100% accuracy, explaining exactly how it reached that conclusion.

2. The "Toolbox Manager" (Anomaly Detection)

• The Old Way: A data analyst has to write code, install software, and manually run complex math programs to find weird glitches in the data. If they pick the wrong tool, they miss the problem.
• The SpaceHMchat Way: You ask, "Is anything weird happening?" The AI acts like a Master Mechanic with a Magic Toolbox. It knows exactly which tool (algorithm) to grab from its shelf to fix the specific problem. It picks the tool, runs the test, and shows you the results. You don't need to know how to code; you just need to know what question to ask.

3. The "Veteran Detective" (Fault Localization)

• The Old Way: When a glitch is found, a senior expert has to spend days studying past cases to guess what broke. "Hmm, this looks like the 2018 battery failure..."
• The SpaceHMchat Way: The AI has been trained on a massive library of past failures. It's like a detective who has read every single police report from the last 50 years. When you show it the data, it instantly says, "This is a broken wire in Battery Pack 2. I know this because the voltage dropped exactly like it did in Case #402." It learns from experience so fast that it can spot common problems better than a tired human.

4. The "Librarian & Strategist" (Maintenance Decision-Making)

• The Old Way: If the satellite breaks, the team has to dig through dusty filing cabinets (digital or physical) filled with design manuals, accident reports, and technical papers to figure out how to fix it. This takes hours or days.
• The SpaceHMchat Way: The AI is a Super-Librarian. You ask, "How do we fix a broken wire?" It instantly scans millions of pages of documents, finds the three most relevant repair guides, summarizes them, and suggests a plan: "Try restarting the system. If that fails, switch to the backup wire. Here is a similar case from 2021 where this worked." It does in 3 minutes what would take a human team 30 minutes.

Why is this a big deal?

• It levels the playing field: A junior assistant can now do the work of a senior expert by talking to the AI.
• It saves money: You don't need to hire thousands of experts; you just need a few experts to manage the AI.
• It's transparent: Unlike some "black box" AI, this system shows its work. It says, "I thought Step 1, then Step 2, so the answer is X." This is crucial for space, where mistakes are dangerous.
• It's open: The creators didn't hide their work. They released the code, the data, and the "training school" (the dataset) for anyone to use.

In a nutshell:
The paper argues that as we fill the sky with thousands of satellites, we can't rely on humans doing everything manually. We need a smart partner (SpaceHMchat) that handles the data, the math, and the paperwork, allowing humans to stay in the loop as the ultimate decision-makers. It's not about replacing the pilot; it's about giving the pilot a co-pilot that never sleeps and knows every manual by heart.

Technical Summary

Here is a detailed technical summary of the paper "Empowering All-in-Loop Health Management of Spacecraft Power System in the Mega-Constellation Era via Human-AI Collaboration."

1. Problem Statement

The rapid deployment of Satellite Mega-Constellations (SMC) (e.g., Starlink, GW, OneWeb) is ushering in an era where the number of spacecraft will increase exponentially. This scale introduces critical challenges for the Spacecraft Power System (SPS), which is the most frequent source of anomalies (accounting for ~44% of cases) and a primary cause of mission failure.

The conventional All-in-Loop Health Management (AIL HM) paradigm faces three insurmountable bottlenecks in the SMC era:

1. Manpower Cost Explosion: Maintaining hundreds of 24/7 expert teams for thousands of satellites is economically and logistically unfeasible.
2. Textual Information Deluge: Critical knowledge resides in vast, unstructured documents (manuals, logs, design specs) that are inefficient for humans to review manually.
3. Operational Complexity: The sheer volume of logical rules, indicators, and fault patterns exceeds human cognitive capacity, leading to slow diagnosis times.

While Large Language Models (LLMs) show promise in other fields, the high-reliability and non-transferable responsibility requirements of aerospace prevent fully autonomous agents. Therefore, a Human-AI Collaboration (HAIC) framework is required to augment, not replace, human operators.

2. Methodology: The AUC Principle and SpaceHMchat

The authors propose the Aligning Underlying Capabilities (AUC) principle, which maps the intrinsic nature of specific health management sub-tasks to the corresponding strengths of LLMs. Based on this, they developed SpaceHMchat, an open-source HAIC framework.

The framework addresses the four stages of AIL HM as follows:

A. Work Condition Recognition (Logical Reasoning Task)

• Human Capability: Logical reasoning and judgment based on decision trees.
• LLM Alignment: Chain-of-Thought (CoT) and Step-wise Prompting.
• Implementation: Human experts pre-define decision trees. These are converted into structured prompts guiding the LLM to reason step-by-step (e.g., "If Voltage > 1V, then..."). The LLM calculates indicators and outputs the work condition with a transparent reasoning chain.

B. Anomaly Detection (Tool-Dependent Task)

• Human Capability: Tool operation and algorithm selection.
• LLM Alignment: Function Calling and Model Context Protocol (MCP).
• Implementation: The LLM acts as an interface to a toolkit containing anomaly detection algorithms (e.g., MTGNN, Transformer, GRU). It interprets user intent, selects the appropriate algorithm, configures parameters, executes the code, and visualizes results, eliminating the need for operators to write code.

C. Fault Localization (Learning-and-Approximation Task)

• Human Capability: Experience accumulation and pattern recognition from historical data.
• LLM Alignment: Fine-tuning (SFT, LoRA, GRPO).
• Implementation: Time-series telemetry data is converted into a Question-Answer (Q&A) dataset. The base LLM is fine-tuned to become an "expert" capable of identifying fault types (e.g., open circuits, short circuits) and providing the reasoning behind the diagnosis, mimicking senior subsystem experts.

D. Maintenance Decision-Making (Knowledge-Intensive Task)

• Human Capability: Searching, summarizing, and synthesizing vast technical documentation.
• LLM Alignment: Retrieval-Augmented Generation (RAG) and Knowledge Bases.
• Implementation: A vector database is built from historical root cause reports, design docs, and accident case studies. When a fault occurs, the LLM retrieves relevant top-k passages, synthesizes root cause analyses, assesses risks, and proposes maintenance strategies (e.g., redundancy activation, reboot) with cited references.

3. Experimental Validation & Dataset

To validate the framework, the authors created a comprehensive experimental ecosystem:

• Hardware-Realistic Platform: A physical testbed fully replicating a real SPS, capable of fault injection.
• Simulation Model: A physics-informed simulation model open-sourced for digital twin research.
• XJTU-SPS Dataset: The first publicly available AIL HM dataset for SPS, containing:
  • 4 Sub-datasets: Work Mode Recognition, Anomaly Detection, Fault Localization, and Forecasting/Reconstruction.
  • Scale: Over 700,000 timestamps, 17 fault types, 33 sensors, and 6 working conditions.
  • Resources: Available via GitHub and Hugging Face.

4. Key Results

SpaceHMchat was evaluated against 23 quantitative metrics, demonstrating superior performance compared to base LLMs without these specific techniques:

Module	Key Metric	Performance
Work Condition Recognition	Conclusion Accuracy	100% (50/50 tasks)
	Step-wise Correctness	100% (350/350 steps)
	Reasoning Time	< 3 seconds
Anomaly Detection	Tool Invocation Success Rate	>99% (50/50 tasks)
	Tool Invocation Accuracy	99.6% (250/251 calls)
	Algorithm Selection Time	< 1 second
Fault Localization	Precision	>90%
	F1-Score	89.35%
	Localization Time	< 1 minute
Maintenance Decision	Search & Analysis Time	< 3 minutes (vs. >30 mins manually)
	Citation Accuracy	97.3% (939/965 valid refs)
	Task Completion Rate	100% (Helpful guidance provided)

Comparative Advantage:

• Base Models (e.g., Qwen3-235B) failed to execute local tools, hallucinated reasoning steps, or lacked specific domain knowledge without fine-tuning/RAG.
• SpaceHMchat successfully integrated tools, provided transparent reasoning chains, and accessed specific domain knowledge, enabling junior assistants to perform tasks previously requiring senior experts.

5. Key Contributions

1. The AUC Principle: A novel design philosophy for HAIC that aligns specific LLM capabilities (reasoning, tool use, learning, knowledge retrieval) with the intrinsic nature of health management sub-tasks.
2. SpaceHMchat Framework: An open-source, conversational AI copilot that unifies the entire AIL HM loop, reducing the barrier for junior staff and optimizing personnel structure (experts focus on iteration and complex cases).
3. XJTU-SPS Dataset & Platform: The creation of the first open-source, hardware-realistic dataset and simulation model for SPS health management, addressing the scarcity of public aerospace data.
4. Empirical Validation: Comprehensive demonstration that LLMs, when guided by HAIC, can achieve high accuracy and efficiency in critical aerospace safety tasks.

6. Significance and Future Outlook

• Paradigm Shift: This work shifts the focus from purely algorithmic time-series analysis to a holistic approach that leverages textual knowledge and human-AI collaboration.
• Scalability: It offers a viable solution for managing the upcoming wave of mega-constellations by drastically reducing the need for massive expert teams.
• Transparency: By enforcing CoT and RAG, the system ensures that AI decisions are interpretable and traceable, a critical requirement for aerospace safety.
• Future Direction: The authors envision a transition from the AUC principle (Aligning Underlying Capabilities) to a SUC principle (Surpassing Underlying Capabilities), where LLMs eventually evolve into fully autonomous agents capable of self-improvement and direct command execution, provided responsibility attribution issues are resolved.

In conclusion, this paper establishes that Human-AI Collaboration is the necessary and effective path to managing the complexity of spacecraft power systems in the mega-constellation era, bridging the gap between massive data volumes and limited human cognitive resources.