Automatic Identification Technology for Acoustic Emission Damage Types of Composite Materials that Integrates Unsupervised Clustering and LLM Statistical Profiling

Abstract

To address the challenges of easily confused weak boundary mechanisms and the difficulty in accurately and automatically identifying damage in composite materials under cumulative damage acoustic emission signal classification, this paper employs UMAP dimensionality reduction on the extracted time-frequency-wavelet packet high-dimensional feature set, utilizes the HDBSCAN algorithm for density clustering and noise removal, then performs soft probabilistic modeling for each cluster using a Gaussian mixture model (GMM), and extracts cluster-level statistical profiles containing higher-order moments and correlation matrices based on posterior weights. By constructing rule-based prompts using structured text input, a frozen-parameter LLaMA-3.1 model is driven to perform zero-shot physical chaining inference. Combined with Mahalanobis distance conflict resolution, unsupervised and accurate mapping of damage semantics is achieved. This method achieves an average F1-score of 0.8912 on five damage types: matrix cracking, debonding, pull-out, fiber fracture, and delamination. This technique enables high-precision automatic identification of acoustic emission damage mechanisms in composite materials under unlabeled conditions.