Data fusion for full-range response reconstruction via diffusion models

Wingho Feng, Quanwang Li, Chen Wang, Jiansheng Fan

TL;DR

This paper introduces a novel data fusion framework using diffusion models to reconstruct the complete structural response of buildings from limited sensor data. By leveraging Diffusion Posterior Sampling (DPS) and three flexible forward models, the approach accurately predicts full-range structural behavior, offering a significant advancement for structural health monitoring.

What Problem Does This Paper Address?

Structural health monitoring (SHM) relies on understanding a structure's full-range response to ensure safety. However, practical limitations like cost, accessibility, and scale often restrict the number of sensors that can be deployed. This sparse and sometimes heterogeneous sensor data is insufficient for comprehensive monitoring, leading to ambiguity in assessing structural integrity and potentially missing critical damage. Existing methods, while improving, struggle with nonlinear behavior, computational demands, and effectively integrating sparse, diverse sensor inputs for full-range reconstruction.

Methods

The core of this work is a data fusion framework built upon diffusion models. These generative models learn to reverse a process of gradually adding noise to data, allowing them to generate realistic samples. The key innovation is the integration of Diffusion Posterior Sampling (DPS), which uses sensor measurements as probabilistic constraints to guide the diffusion model's reverse process.

Three novel forward models are proposed to map structural responses to sensor measurements, offering flexibility:

  • Direct Observation Mapping (DOM): The simplest approach, using a selection matrix to extract sensor values directly from the full response. It's efficient but limited to cases where sensor and response types are identical.
  • Channel-based Observation Mapping (COM): Extends DOM by treating different sensor types (e.g., strain and stress) as distinct channels, enabling reconstruction of one type from another. This is useful for multi-physics monitoring.
  • Neural Network Forward Model (NNFM): Employs a differentiable neural network to learn complex, nonlinear mappings between sensor data and structural responses. This offers the most flexibility for heterogeneous data and complex relationships, overcoming limitations of fixed physical models.

Datasets

The proposed framework is validated on data from a steel plate shear wall (SPSW) experiment, specifically the FSPW2 benchmark experiment. The study obtained element-wise response (strain and stress) histories from finite element analysis simulating displacement loading up to 100 mm, ensuring nonlinear behavior was captured. This dataset was used to reconstruct the full-range von Mises stress response.

Key Insights and Ideas

  • Diffusion models are powerful for inverse problems: By framing response reconstruction as a generative task, diffusion models can effectively handle the inherent uncertainty and ill-posed nature of inferring global behavior from local measurements.
  • DPS bridges the gap: Diffusion Posterior Sampling is crucial for incorporating sparse and heterogeneous sensor data into the diffusion process, guiding the generation of realistic full-range responses.
  • Forward model flexibility is key: The choice of forward model (DOM, COM, NNFM) significantly impacts performance and applicability, allowing the framework to adapt to various sensing scenarios and data types.
  • Spatial dispersion matters: Sensor placement strategies that offer broader spatial coverage (standard and random) generally lead to better reconstruction accuracy than concentrated or low-variance placements.
  • Hyperparameter tuning is critical: Parameters like sampling steps and conditional sampling strength significantly influence accuracy, especially for complex models like NNFM, highlighting the need for systematic tuning.

Key Results

  • The framework was validated on a steel plate shear wall exhibiting nonlinear responses.
  • The three forward models achieved good reconstruction accuracy, with Weighted Mean Absolute Percentage Errors (WMAPE) of 1.62% (DOM), 3.27% (COM), and 3.49% (NNFM).
  • DOM, despite its simplicity, achieved the best performance, likely due to its direct task of reconstructing stress from stress measurements.
  • The framework demonstrated robustness to varying hyperparameters, sensor configurations, and noise levels.
  • Inference time was found to be low, with reconstructions of 100 samples completed in under one second under balanced settings, making it suitable for real-time applications.

Limitations and Future Work

  • Hyperparameter Sensitivity: Complex models like NNFM exhibit higher sensitivity to hyperparameter settings, requiring careful and extensive tuning for optimal accuracy.
  • NNFM's Spatial Encoding: The NNFM's MLP component, while powerful, doesn't explicitly encode spatial structure, potentially leading to less precise local modeling compared to physics-based approaches in certain regions.
  • Future Work: The paper suggests that the framework opens new possibilities for probabilistic modeling and sensor data fusion in SHM. The flexibility in modifying forward models and adjusting hyperparameters allows for rapid assessment of alternative sensor placement strategies, offering potential for optimizing sensor deployment. The ability to handle sparse and noisy data paves the way for real-time structural estimation and enhanced infrastructure safety.

Summary updated

BibTeX
@article{feng2026data,
  author = {Wingho Feng and Quanwang Li and Chen Wang and Jiansheng Fan},
  title = {{Data fusion for full-range response reconstruction via diffusion models}},
  journal = {Adv. Eng. Informatics},
  year = {2026},
  doi = {10.1016/j.aei.2026.104499},
}
Figure