Technical Summary: SHM-Agents

Problem Statement

In the field of Structural Health Monitoring (SHM), existing specialized algorithms are effective but suffer from high implementation barriers, limited interoperability, and complex training procedures. Users typically face fragmented workflows where they must understand and operate multiple independent algorithmic modules with incompatible interfaces. This necessitates significant time investment in data preparation, model training, inference deployment, and manual workflow coordination. Furthermore, while Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding, their inherent limitation of producing predominantly text-based outputs prevents them from independently executing complex engineering tasks. Current SHM systems lack an integrated solution that can seamlessly bridge natural language commands with specialized algorithmic execution.

Methodology

To address these challenges, the authors propose SHM-Agents, a generalist-specialist integrated agent system. This system leverages the reasoning and planning capabilities of general LLMs while integrating the problem-solving strengths of specialized SHM algorithms.

System Architecture

The architecture is modular, consisting of three primary components:

1. Process Agent Nodes: These nodes orchestrate the workflow.
   • Architecture Node: Generates task execution plans based on user input and algorithmic details.
   • Allocate Node: Evaluates the plan, formulates specific commands for skill nodes, monitors execution, and handles error correction via a "allocate-execute-feedback-reallocate" mechanism.
   • Summary Node: Encapsulates dialogue, summarizes execution flows and outcomes, and delivers final feedback to the user.
2. Skill Agent Nodes: These nodes perform specific SHM tasks using dedicated algorithms. The system currently includes eight specialized nodes (e.g., Data Anomaly Diagnosis, Modal Identification, Vehicle Load Modeling, Fatigue Calculation, Reliability Calculation, RAG for Q&A) and an "Other Node" capable of writing and executing Python code for custom tasks like signal processing or distribution fitting.
3. Global Modules: These manage system state and data integrity.
   • Algorithm Module: Stores input/output specifications of specialized algorithms to ensure accurate invocation by the LLM.
   • Data Module: Manages variables generated during execution, ensuring accurate data transfer and format transformation between nodes.
   • Config Module: Handles bridge-specific configurations, including finite element models, optimal parameters for deep learning algorithms, and database details. It supports automated pre-training and configuration generation.
   • Memory Module: Stores contextual dialogue information to maintain conversation history and context.

Workflow

The system operates on an "architecture–allocation–execution–summary" workflow. Upon receiving a natural language command (and optional data), the Architecture node creates a plan. The Allocate node breaks this down into subtasks, assigning them to appropriate Skill nodes. If an error occurs, the Allocate node attempts to regenerate commands; if unresolvable, it reports to the user. Finally, the Summary node compiles the results.

Key Contributions

The paper outlines four primary contributions:

1. System Architecture: The development of SHM-Agents, which integrates an LLM as an intelligent core with specialized SHM algorithms, demonstrating strong capabilities in organizing and executing complex tasks via a self-feedback regulation mechanism.
2. End-to-End Execution: The ability to execute single or combined SHM tasks through natural language interaction. The system supports deep model pre-training without human intervention, significantly simplifying the deployment of specialized algorithms.
3. Modular Management: The introduction of an Algorithm Module to manage diverse algorithms and a Data Module to ensure accurate variable transfer, facilitating the correct invocation of tools.
4. Scalability: The system is designed for high scalability, allowing users to rapidly add new agent nodes to address additional tasks without overhauling the core system.

Experimental Results

The system was evaluated on a long-span cable-stayed bridge in China. The experiments demonstrated the system's ability to accurately and efficiently perform a wide range of tasks, both individually and in combination:

• Data Anomaly Diagnosis and Recovery: Successfully detected anomalies (e.g., missing data, outliers) and reconstructed data using alignment information from adjacent time periods.
• Signal Processing: The "Other Node" autonomously generated and executed Python code to perform Fourier and wavelet transforms, adhering to complex visualization requirements.
• Modal Identification: Performed structural modal identification on reconstructed data, successfully identifying mode shapes, frequencies, and damping ratios.
• Long-term Analysis: Analyzed long-term variations in modal parameters over time intervals.
• Reliability Assessment: Calculated structural reliability and failure probability for the current state and projected trends over the next 10 years, incorporating corrosion models.
• Fatigue Estimation: Computed fatigue damage based on vehicle loads generated from WIM data.
• Structural Response Calculation: Calculated mid-span displacement under dynamic vehicle loads using PyAnsys.
• Knowledge Q&A: Utilized a Retrieval-Augmented Generation (RAG) node to answer specific questions about bridge materials and specifications based on uploaded documents.

Significance and Claims

The paper claims that SHM-Agents represents a significant step toward streamlining SHM workflows. By integrating LLMs with specialized algorithms, the system reduces the complexity of agent construction and lowers the barrier to entry for using advanced SHM tools. It enables non-experts to perform complex, multi-step engineering analyses through natural language.

The authors note that while the system is highly effective, its problem-solving accuracy is directly correlated with the capabilities of the underlying LLM (e.g., GPT-4o), achieving operational accuracy above 95% but not 100%. Additionally, the ultimate quality of results remains contingent upon the performance of the underlying specialized algorithms. The paper concludes that while some planned algorithms (e.g., visual surface damage detection, digital twin simulation) are not yet implemented, the modular architecture allows for their continuous addition, offering a flexible path for future expansion.