Bone damage detection using ultrasonic guided waves: multi-feature integration and machine learning approaches

Reliable detection of bone damage remains challenging due to the complex propagation of ultrasonic guided waves (UGW) and the variability of biological tissues. Most existing methods rely on single-domain handcrafted features, which limits robustness and generalisation, particularly in small-sample and heterogeneous scenarios. This study proposes a non-invasive bone damage detection framework that integrates multi-domain feature extraction, genetic algorithm (GA)-based feature selection, and a multi-feature input fully connected neural network (MFI-FCNN). UGW signals were collected from bovine cortical bone specimens with three damage severities at varying positions. A 21-dimensional feature set was constructed from time-domain, frequency-domain, and time-frequency descriptors, together with two CEEMD-based features: energy entropy and an average spatial distance-based damage metric. GA-based feature selection enhanced class separability and reduced redundancy. The proposed framework was evaluated under three protocols, including single-position, multi-position, and small-sample classification. Experimental results show that MFI-FCNN consistently outperformed conventional methods, including SVM and shallow MLP, achieving up to 100% accuracy on independent validation sets and exhibiting strong robustness to noise. These results demonstrate the effectiveness and adaptability of the proposed framework for UGW-based bone damage detection.