SWT-ResNet-based internal material identification for impacted multilayered cylindrical structures with clearance

Abstract Accurate identification of materials enclosed within multilayer cylindrical structures is challenging without structural disassembly, particularly when assembly clearances induce contact and impact nonlinearities and result in nonstationary vibration responses. To address this problem, a material identification framework integrating the synchrosqueezing wavelet transform and a residual neural network (SWT–ResNet) is proposed. First, vibration responses are measured externally through impact tests, and SWT is employed to generate highly concentrated time–frequency representations through frequency reassignment. A lightweight ResNet is then used to automatically extract multilevel time–frequency features and classify the internal materials. Four materials, namely MJ, AZ31B magnesium alloy, epoxy resin, and polytetrafluoroethylene, are investigated under four impact levels of 5, 10, 15, and 20 g. Three independent repeated tests are performed for each condition, producing 48 original experimental records. To prevent information leakage caused by overlapping sliding windows, the data are partitioned at the original-record level before sample segmentation. The proposed SWT–ResNet achieves an accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1-score of 97.50%, 97.56%, 97.50%, and 97.50%, respectively, on the independent test set. Original-record-level three-fold cross-validation yields an accuracy of 96.88% ± 0.63%, demonstrating the stability of the proposed framework. Ablation experiments show that SWT–ResNet outperforms CWT–ResNet, STFT–ResNet, Raw-CNN, SWT–CNN, SWT–SVM, SWT–RF, and PCA–k-means, confirming the contributions of both the highly concentrated SWT representation and residual feature learning. Under Gaussian white-noise conditions of 20, 10, and 0 dB, the corresponding accuracies are 98.5%, 94.4%, and 69.3%, respectively, indicating a certain degree of noise robustness. These results demonstrate that SWT–ResNet can effectively identify internal materials in multilayer cylindrical structures with assembly clearances and provides a feasible nondestructive approach for identifying internal media in complex enclosed structures.