| name | llm-agentic-fault-tolerant-control |
| description | Agentic Large Language Model framework for active Fault-Tolerant Control (FTC) in cyber-physical systems. Combines multi-agent LLM workflow with Digital Process Plant Twin and Graph RAG (CPSMod ontology) to transform fault detection outputs into constraint-aware recovery actions. Suitable for industrial control systems, process automation, and safety-critical applications.
|
| tags | ["fault-tolerant-control","llm-agents","cyber-physical-systems","graph-rag","digital-twin","process-control"] |
| created | 2026-06-30T00:00:00.000Z |
| source | arXiv:2606.28011 |
LLM Agentic Fault-Tolerant Control (FTC)
Core Contribution
This paper presents an agentic LLM framework that transforms fault detection outputs into constraint-aware recovery actions for industrial control systems. The approach bridges the gap between fault detection and validated corrective action through a multi-agent workflow.
Architecture Components
1. Multi-Agent Workflow
The framework decomposes operator duties into specialized agents:
- Monitoring Agent: Observes system state and fault signals
- Planning Agent: Generates recovery strategies
- Action Synthesis Agent: Creates specific control commands
- Simulation Agent: Tests actions in Digital Twin before execution
- Validation Agent: Verifies actions against safety constraints
- Reprompting Agent: Refines actions based on feedback
2. Digital Process Plant Twin (DPPT)
Exposes plant capabilities through:
- Real-time plant data access
- Process models for simulation
- Simulation service for pre-execution testing
- Interlock and envelope validation
3. Graph RAG Layer (CPSMod Ontology)
Organizes plant knowledge into a graph structure:
- Structure: Physical components and relationships
- Function: Process functions and dependencies
- Hybrid Dynamics: Continuous/discrete state transitions
- Control Context: Operating modes and constraints
- Fault Semantics: Fault types and propagation paths
Supports relation-aware, multi-hop retrieval for agents.
Recovery Action Generation
Action Types
- State-machine recovery paths: Minimal-risk sequences of state transitions
- Discrete commands: Valve operations, mode switches, equipment activation
- Continuous setpoint adaptations: Adjusted control targets
Validation Pipeline
Before any actuation:
- Interlock check: Verify physical safety constraints
- Envelope validation: Ensure within operational limits
- Dynamic feasibility: Confirm temporal and rate constraints
- Fallback trigger: If no valid plan found, hand over to safety system
Implementation Pattern
class FTCFramework:
def __init__(self, dppt, graph_rag, llm_agents):
self.dppt = dppt
self.graph_rag = graph_rag
self.agents = llm_agents
def handle_fault(self, fault_signal):
diagnosis = self.agents.monitor(fault_signal)
strategy = self.agents.plan(diagnosis, self.graph_rag.query())
actions = self.agents.synthesize(strategy, self.dppt.get_constraints())
for action in actions:
if self.dppt.simulate(action).is_safe():
return action
return self.agents.fallback(diagnosis)
Performance Characteristics
- Latency: Compatible with process dynamics (batch: minutes, continuous: seconds)
- LLM Models: Tested with lightweight models (GPT-4o-mini, GPT-4.1-mini)
- Validation: Deterministic constraint checking ensures safety
- Applicability: Both discrete (batch) and continuous (CSTR) processes
Use Cases
- Chemical Process Control: Reactor temperature/pressure fault recovery
- Batch Manufacturing: Mixing module fault handling
- Power Systems: Grid fault detection and reconfiguration
- Manufacturing: Production line fault recovery
- Building Automation: HVAC system fault management
Integration Requirements
- Graph Database: For CPSMod ontology storage and querying
- Digital Twin Platform: Real-time simulation capability
- LLM API: For agent reasoning and decision making
- Control System Interface: For reading state and executing actions
- Safety System: For fallback and emergency shutdown
Limitations and Considerations
- Requires comprehensive plant knowledge in CPSMod ontology
- LLM latency must be compatible with process dynamics
- Digital Twin accuracy affects validation reliability
- Graph RAG quality depends on ontology completeness
- Safety fallback must be independent of LLM decisions
References
- Paper: arXiv:2606.28011v1
- Authors: Javal Vyas, Milapji Singh Gill, Artan Markaj, Felix Gehlhoff, Mehmet Mercangöz
- Published: 2026-06-26
- Categories: eess.SY, cs.LG