Generation of Fault Data from Multiple Types of Bridge Monitoring Sensors Based on Time Series Diffusion Models Fusing Control Conditions and Pseudo Prompt Enhancement

Bridges are crucial infrastructure components in transportation networks. However, as they age and traffic loads increase, their safety faces significant challenges. Bridge health monitoring systems collect structural data through sensors. However, obtaining sensor failure data is difficult, and traditional data generation methods, such as manually adding standard deviations, often lack data diversity. As a result, they fail to effectively reflect realistic sensor failures, which impairs the generalization ability and accuracy of fault diagnosis models. This paper proposes a bridge sensor fault data generation method based on a time series diffusion model, combining control conditions with pseudo prompt enhancement techniques. The goal is to improve the diversity and quality of the generated data. First, the Variational Autoencoder (VAE) is jointly trained with the inverse denoising process of the diffusion model to generate structured noise and enhance the complex characteristics of the noise. Then, a control condition module is introduced to regulate the quality of noise generation. To address the issue of insufficient data samples or uncertainty in labeling, a Pseudo Prompt Enhancement module is proposed, which utilizes a pre-trained autoencoder or self-supervised learning method to generate pseudo prompts that provide auxiliary information about the sensor device status. Furthermore, a classifier-free guidance mechanism is incorporated into the model training process to further enhance the quality and diversity of the generated data. Experimental results demonstrate that the proposed method yields significant improvements in generating real bridge sensor data. This approach offers a promising solution for generating realistic sensor fault data, advancing bridge health monitoring systems and enhancing their diagnostic accuracy.