Abstract:

To achieve high-precision recognition of wall hollowing sound signals and improve multi-category detection accuracy, a multi-feature fusion method for wall hollowing detection based on BO-SVM was proposed. First, the collected knocking sound signals from different wall types were preprocessed by pre-emphasis, framing, and windowing, and both MFCC and MSC were extracted. The two acoustic features were concatenated at the frame level and normalized to construct a fused feature dataset. Then, a BO-SVM classification model was developed, and the kernel function penalty and parameters were optimized using five-fold cross-validation to establish the MFCC+MSC-BO-SVM model. Finally, classification experiments were conducted using hollow and non-hollow from multiple wall types, including cement, coating, marble, and ceramic tile walls. The results show that the fused features outperform single features in terms of accuracy, recall, and F₁-score. The MFCC+MSC-BO-SVM model achieves an overall recognition accuracy of 96.36%, representing improvements of 6.61%, 9.58%, 15.27%, 13.90%, and 5.02% compared with standard SVM, Random Forest, K-Nearest Neighbor, Grid Search-optimized SVM, and Chaos Particle Swarm Optimization SVM respectively. Furthermore, the BO method can obtain the optimal parameter combination with fewer iterations, showing superior convergence and classification stability.

Key words: feature fusion, Bayesian optimization (BO), support vector machine (SVM), wall hollow drum detection, Mel-frequency cepstral coefficients (MFCC), Mel-spectral coefficients (MSC)