A 1DCNN-SENet-LSTM model for UGW-based automatic evaluation of grouting void location in post-tensioned concrete structure

Abstract

Void is a common grouting defect that significantly impacts the safety and durability of post-tensioned concrete structures. However, due to invisibility, it is challenging to detect and evaluate grouting voids. This study proposes a one-dimensional convolutional neural network-squeeze and excitation net-long short-term memory (1DCNN-SENet-LSTM) model for UGW-based automatic evaluation of grouting void location. In this model, the UGW time-domain signal is filtered and input into the network. Firstly, CNN is employed to extract spatial features. Secondly, the SE block is incorporated for attention calibration. Finally, the LSTM is utilized to extract time series-related features. The implementation of 9-fold cross-validation during network training enhances robustness and avoids overfitting. The proposed employs an end-to-end approach to facilitate automatic feature learning. This study compares the test results of the 1DCNN-SENet-LSTM model with other models to validate the superiority of the proposed model in terms of prediction performance. The classification results are evaluated using indicators such as Accuracy, Precision, Recall, and F₁ score. The results indicate that the 1DCNN-SENet-LSTM model achieves a greater Accuracy of 96 % in evaluating the location of grouting voids, which is 10.3 % higher than the 1DCNN-LSTM model, 12.9 % higher than the ResNet model, and 23.1 % higher than the XGBoost model. Furthermore, the F₁ score for each category in the 1DCNN-SENet-LSTM model exceeds 92 %, surpassing other models.