Robust composite damage detection via feature subspace ensemble learning with data distribution shifts

Machine learning has greatly improved damage detection in composite structures, with models like random forests and support vector machines as key examples. However, single-model methods often struggle with low accuracy and inconsistency due to real-world challenges such as noise and shifting data patterns. To tackle these problems, this study introduces a new ensemble learning approach based on feature subspaces to enhance damage detection under varying data conditions. The method first uses an improved variational mode decomposition algorithm to filter out noise and identify key patterns in vibration signals collected from different measurement points, which naturally show varied data distributions. Then, it extracts and evaluates multiple features to build diverse subspaces that capture signal changes. Finally, a stacked ensemble technique combines models trained on these subspaces to deliver precise and stable damage detection. The method is tested on a custom-built platform for inflatable composite structures, focusing on fibre fracture damage. Experimental results show that this method outperforms traditional techniques, especially in cases where data patterns shift.