Asynchronous Multimodal Sensor Data Fusion for Label-Scarce Health Indicator Mining in Composite Structure Monitoring
Main article
Abstract
Structural health monitoring (SHM) of composite aerospace panels requires comprehensive health indicators (HIs) that capture multiscale damage evolution under complex fatigue loading. However, constructing reliable HIs is hampered by the asynchronous acquisition rates of passive and active sensing modalities, the scarcity of ground-truth labels, and the heterogeneous information content of complementary sensor streams. This study proposes a multi-level asynchronous multimodal sensor data fusion framework that integrates passive acoustic emission (AE) and active guided wave (GW) sensing to mine high-quality HIs for T-stiffener composite panels without labeled failure data. Three intra-modality AE pipelines and one GW pipeline are developed under an inductive semi-supervised learning paradigm, using physics-informed proxy labels and criteria-driven regularization to enforce monotonicity, prognosability, and trendability. A causal zero-order-hold synchronization scheme aligns the dense AE timeline with sparse GW acquisitions. Late-fusion regression models—including Gaussian process regression, gradient boosting, support vector machines, multilayer perceptrons, and long short-term memory networks—are benchmarked for inter-modality HI fusion. Leave-one-out cross-validation on 12 AE units and 5 GW units demonstrates that the fused HIs consistently exceed a Fitness score of 90%, with the LSTM-based fusion achieving up to 97% ± 2% cohort-level performance. The proposed framework offers a scalable and deployable blueprint for next-generation prognostics in safety-critical composite structures.
