Data augmentation of dynamic responses for structural health monitoring using denoising diffusion probabilistic models

In the field of structural health monitoring, deep learning techniques are gaining increasing recognition, with the fundamental requirement of high-quality data for effective implementation. This paper addresses the challenges related to data imbalance and inadequacy by proposing the development and application of diffusion models for data augmentation. When training diffusion models, the model learns to predict the noise introduced into an initially clean measurement signal. During the generation of signal samples, the diffusion model starts with Gaussian noise and progressively denoises it to generate realistic samples. Adaptations to the training framework and meticulous design of the model architecture are implemented to enhance their suitability for generating signals in the context of structural health monitoring. The authenticity of the generated signals is validated by two case studies. The first one consists of 7-channel samples generated to simulate the acceleration responses of a high-rise building under ambient excitations. Modal identification is executed using a generated sample, yielding modal parameters consistent with those from previous studies. The second one generates 16-channel acceleration responses of a suspension bridge excited by a hurricane for data augmentation. The authenticity is confirmed through a lost response reconstruction task, where reconstruction accuracy is enhanced when the generated signals are used as supplementary training data. The results from experimental studies demonstrate the realistic time-spatial dependencies inherent in the generated data, and the accuracy and effectiveness of using diffusion models to generate structural response datasets for SHM applications. • This paper proposes a novel approach for data augmentation of dynamic responses. • The proposed approach is built on denoising diffusion probabilistic models. • The effectiveness and accuracy of data augmentation are demonstrated on two case studies. • Modal identification is conducted to demonstrate the validity of the generated data. • Generated data are used as supplementary training data for lost response reconstruction.