Robust structural damage identification through VMD and feature-level fusion of multi-source vibration signals

Structural damage identification using deep learning is often constrained by a reliance on single-data-type models and the challenges posed by non-stationary signals. This paper introduces a novel framework that integrates Variational Mode Decomposition (VMD) with multi-source data fusion to overcome these limitations. The proposed method processes raw, non-stationary acceleration and strain signals through VMD to generate stable intrinsic mode functions (IMFs), significantly enhancing their feature clarity for a subsequent convolutional neural network (CNN). A comparative analysis on the Vänersborg and KW51 bridge datasets confirms that VMD preprocessing consistently improves classification accuracy for both data types compared to using unprocessed signals. The core finding demonstrates that feature-level fusion of these VMD-processed signals is the optimal strategy, achieving peak accuracies of 98.4% and 96.6% on the Vänersborg and KW51 datasets, respectively. This approach not only surpassed the performance of acceleration-only models but also substantially bridged a critical performance gap by boosting the contribution of strain data by over 23% and improving strain-only model accuracy by 24.5%. While data-level and decision-level fusion showed context-dependent results, feature-level fusion proved uniquely robust and superior across both structures. The study conclusively establishes that the feature-level fusion of VMD-decomposed signals is a highly effective and reliable methodology for advanced structural health monitoring. To further support the effectiveness of VMD, it is additionally benchmarked against conventional baseline decomposition techniques Empirical Mode Decomposition (EMD) and Ensemble Empirical Mode Decomposition (EEMD), and the results confirm that VMD consistently outperforms both methods under single-signal (data-only) evaluation as well as under feature-level fusion, demonstrating its superior decomposition capability and practical robustness for damage identification.