Comparative analysis of data-driven autoencoder networks for full-field expansion from sparse measurements

Condition monitoring relies on data collected from a system with sensors to extract information about its state. The difficulty in deploying highly dense distributions of sensors hinders the detection of local damage in the system. To address this issue, domain-specific techniques have been developed that can expand measurements from a discrete subset of data (i.e., sparse measurements) to full-field. However, these domain-specific expansion techniques have shown limitations when used with non-linear dynamic systems. Recent advancements in machine learning algorithms, particularly autoencoder (AE) networks, can improve the robustness of expansion techniques to non-linear systems as well as generalize their applicability to other domains. In this research, three AE architectures, based on feed-forward networks, convolutional neural networks, and long short-term memory (LSTM) networks, are proposed to reconstruct the full-field response of a targeted dynamic system when only sparse measurements are available. The performance of the three architectures is compared as a function of i) the location of measurement points, ii) the spatial density of the measurement points, and iii) the noise present in the signal collected at each measurement point. Tests performed on analytical and experimental datasets indicate that all three architectures successfully expanded to full-field data from sparse measurements, with LSTM exhibiting errors below 0.85% in the expansion process. Advancements of this research could lead to the application of AE-based expansion techniques for condition monitoring of dynamic systems when only a limited number of measurements are available.