Abstract
To address the limitations of the conventional low-strain reflected wave method for pile foundation testing, this study proposes an intelligent multi-point vibration acquisition testing model based on machine learning to evaluate the integrity of in-service, high-cap pile foundations. The model's performance was assessed using statistical error metrics, including the correlation coefficient R(2), mean absolute error (MAE), and variance accounted for (VAF), with comparative evaluations conducted across different model frameworks. Results show that both the convolutional neural network (CNN) and the long short-term memory neural network (LSTM) consistently achieved high accuracy in identifying the location of the first reflection point in the pile shaft, with R(2) values greater than 0.98, MAE below 0.41 (m), and VAF greater than 98%. These findings demonstrate the model's strong predictive capability, test stability, and practical utility in supporting operator decision-making. Among the evaluated models, CNN is recommended for analyzing the integrity of in-service pile foundation based on the multi-point vibration pickup signals and multi-sensor fusion signal preprocessed by the time series stacking method.