Regression prediction of bone damage using triple-branch convolutional neural network based on ultrasonic guided waves

Osteoporosis frequently causes occult bone injuries and fractures, creating a demand for efficient, non-invasive, and intelligent bone health assessment. This study proposes a triple-branch convolutional neural network (TB-CNN) that integrates multiscale features to perform quantitative regression modelling of bone damage using ultrasonic guided wave signals. An in vitro bovine bone dataset containing eight damage depths (0–7 mm) was constructed, and the training data were preprocessed with Z-score normalisation and augmented by Gaussian noise and amplitude scaling. The proposed model extracts deep features from the time domain, frequency domain, and six types of time – frequency-domain transformation maps (e.g. Wigner – Ville distribution, continuous wavelet transform, and synchrosqueezing transform) and applies a multimodal fusion module for integrated modelling. On the test set, TB-CNN achieves a root mean square error of 0.0053 and a coefficient of determination of 0.9998, outperforming models such as ResNet, support vector regression, multilayer perceptrons, and linear regression. Without relying on manual feature design, the proposed model demonstrates a strong nonlinear modelling ability and predictive accuracy, showing promise for non-destructive bone tissue evaluation.