Disentangling the latent space of an end2end generative VRNN model for structural health condition diagnosis

Despite advancements in deep learning (DL) for Structural Health Monitoring (SHM), interpretability remains a challenging issue. We address this point by introducing physical constraints on the latent space of the DL model to enhance the interpretability on representing the external loads and environment. A generative Variational Recurrent Neural Network (VRNN) model for time-series data from the SHM system of long-span bridges is proposed in this study. Characteristics of the external loads and environment are used to form the constraints on the latent space of VRNN. This approach combines inference and generation processes in the end-to-end DL training, where the encoder in the VRNN model is designed to infer the external loads and environment from the structural response monitoring data and the decoder generates the structural response under the estimated loads and environment, and the reconstruction residual is used as the structural health condition indicator. The pretraining and finetuning framework is employed, that is the simulation data with exact load and environment information is used for pretraining and the real-world output-only data without knowing the actual loads and environment from SHM systems is input in the finetuning. The proposed method is validated on the damage identification benchmark dataset that was released on IPC-SHM2020, and results verify the effectiveness and show that the proposed VRNN model infers the latent space with the physical meaning of vehicle loading with high accuracy in the pretraining stage and can be furtherly applied to the damage diagnosis of real-world applications after finetuning.