AutoML-Optimized Vision Transformers for Multimodal Structural Health Monitoring from Multi-Channel Feature Tensors

Abstract

Reliable structural health monitoring (SHM) of wind turbines requires methods capable of handling high-dimensional, heterogeneous multi-sensor data under both labeled and label-scarce conditions, while remaining computationally efficient for real-world deployment. This paper presents a unified SHM framework that integrates feature-level data fusion, Vision Transformer (ViT) classification, multi-objective AutoML optimization, and unsupervised anomaly detection within a single pipeline. Multi-channel signals from 26 sensors are preprocessed and encoded into a compact three-channel feature tensor combining statistical descriptors, spectral and wavelet features, and a PCA-denoised inter-sensor correlation matrix. This representation enables efficient storage of time-window information and captures both local signal characteristics and global cross-sensor dependencies. A lightweight ViT is trained for supervised fault classification, while its architecture is optimized using NSGA-II to jointly maximize predictive performance and minimize computational cost, enabling deployment on resource-constrained edge devices. To address label scarcity, an autoencoder trained solely on normal-condition data is used for anomaly detection. Comparative evaluation of reconstruction- and latent-space-based metrics shows that Mahalanobis distance in the latent space provides superior sensitivity to subtle faults. Validation on the ETH Aventa AV-7 dataset demonstrates up to 98.4\% macro-F1 in classification and robust anomaly detection performance. The results confirm that the proposed multisensor fusion strategy provides a reliable and scalable pipeline for SHM of real structures, applicable to both supervised and unsupervised scenarios under practical computational constraints.