NDE Data Correlation Using Encoder-Decoder Networks with Wavelet Scalogram Images

Nondestructive evaluation (NDE) technologies are increasingly used to provide information about the subsurface integrity of structural components. While NDE can provide powerful insights, efficiently and consistently correlating information from different NDE sources poses significant challenges. Heterogenous NDE data capture different representations of subsurface phenomenon. This can lead to challenges when trying to associate information from multiple NDE sources across a structure's life cycle. Conceptually, one way to address this is to characterize NDE data in terms of the fundamental (latent) information that is shared between NDE data types, though this remains an understudied problem. An improved understanding of the correlation among the heterogenous NDE data would enable new pathways for the engineers to integrate and analyze NDE data, and potentially lead to a new form of structural health monitoring for multi-NDE systems. This paper presents an approach to identifying these shared latent features through a combination of wavelet signal transformations and an encoder-decoder neural network, or autoencoder. NDE data from two sources is first transformed into 2D scalogram images using a continuous wavelet transform, a process shown to improve subsequent NDE analyses. Then an autoencoder is designed and trained to take one NDE data type as input and reconstruct a representation for the paired NDE method. In the middle of the autoencoder, the model learns a set of reduced-dimension latent features that are shared between paired NDE data types. These autoencoder-generated features can serve as a uniform basis for time-history analyses where NDE methods have varied over a structure's lifespan and can enhance fundamental data interpretation. This process is illustrated on NDE data collected from a series of laboratory-scale tests representing the assessment of concrete bridge decks.