A novel acoustic emission signal analysis method for damage identification in 3D woven composites based on a pre-trained convolutional neural network

Three-dimensional woven composites (3DWC) are extensively utilized across various industries, but their failure mechanisms are complex, and the number of acoustic emission (AE) signals generated during testing is large and complicated. To address the limitations of traditional AE analysis in effectively selecting key damage signals and in identifying damage severity and damage mode, this study proposes a dual-path AE signal analysis method for damage identification in 3DWC. The method integrates multiple time-domain AE features clustering with a pre-trained convolutional neural network (CNN) model. Guided by Pearson correlation coefficient analysis (PCCA) for feature selection, k-means clustering based on amplitude, energy, and root mean square (RMS) screens out key damage signals capturing over 88% of energy. Combined with Fast Fourier Transform (FFT)-based frequency-band energy clustering and time-frequency maps generated by Continuous Wavelet Transform (CWT), the pre-trained CNN model achieves a classification accuracy of 93.88% on the warp-direction tensile dataset. Further validation by test results obtained under four loading conditions demonstrates the ability of the proposed method to identify damage modes.