Multiscale Feature Fusion Convolutional Neural Networks for Low-Speed Impact Localization of Composite Structures Using a Single Sensor

Wind turbine blades and similar composite materials frequently sustain low-velocity impact damage during fabrication and deployment, underscoring the demand for reliable, low-cost localization systems. Here, we present a novel deep learning-based framework that relies on just one sensor to pinpoint impacts on the blade spar. The technique utilizes complete ensemble empirical mode decomposition with adaptive noise to separate non-stationary acoustic emission data into fundamental intrinsic mode functions. A feature-dense training set is generated by reshaping the four IMFs with the maximum correlation values into grayscale matrices. For the source identification phase, a novel MSFF-CBAM network is deployed to automatically extract and optimize multiscale feature representations. Subsequent steel ball drop evaluations indicate that this proposed methodology secures an absolute localization precision of 95.67%. When considering the rationality of boundary errors, the accuracy reaches 99%. This study provides a robust, high-precision solution for structural health monitoring of complex composite structures using minimal sensor hardware.