An intelligent detection method for anchorage-end tension of pre-stressed steel strands based on acoustic signals and MIC-LDA-3CNN
Frequency-domain statistical analysis confirms minimal interference from detailed anchorage configurations on acoustic feature stability. From time-, frequency-, and cepstral-domain features, a joint maximum information coefficient (MIC) and linear discriminant analysis strategy selects 15 sensitive features. A three-layer convolutional neural network (3-CNN) regression model achieves high-precision tension prediction. Experimental results show the 3-CNN outperforms support vector regression, with an overall recognition rate approaching 90% (above 80% in high-tonnage ranges) and a test-set coefficient of determination coefficient of determination = 0.943. Key improvements include non-contact operation (no external excitation beyond lightweight rebound hammer), lightweight portable equipment, strong anti-interference via pinna-coupling, and robustness to anchorage detailing. This method offers a practical, cost-effective solution for rapid on-site detection in existing pre-stressed structures. It can be extended to health monitoring of similar anchorage systems, such as bolts and cables, enhancing structural safety assessment and maintenance decision-making in civil infrastructure.