Computation of the Dynamic Deflection Basin of Asphalt Pavement Based on Convolutional Neural Network, Long Short-Term Memory, and Attention Mechanism

Abstract

Accurately obtaining the dynamic deflection basin (DB) under falling-weight-deflectometer (FWD) loading is a critical indicator for pavement construction quality acceptance and pavement health condition assessment. To provide substantial and reliable data support for pavement construction and maintenance, this study proposes a dynamic DB database construction method based on convolutional neural network, long short-term memory, and attention mechanism (CNN-LSTM-AM). The model uses pavement structural layer parameters and static DB data derived from layered elastic system theory as inputs. Through CNN, local features are efficiently extracted, and the resulting feature vectors are transformed into time-series inputs for the LSTM network model to uncover temporal dependencies. Finally, AM is introduced to dynamically allocate weights to key information, enhancing the model’s ability to capture core features, ultimately constructing a dynamic DB generation model for asphalt pavements. The model demonstrates a relative mean absolute percentage error (MAPE) below 0.09% when predicting dynamic DBs for three-layer and four-layer pavement structures compared with theoretical calculations. Both FWD-measured dynamic DBs and three-dimensional finite element simulation results show high consistency with the computational outcomes based on this method, indicating the model’s excellent accuracy and generalization capability. This provides a new approach for batch acquisition of reliable dynamic DB data, supporting pavement modulus inversion and health assessment, and holds significant practical implications for optimizing maintenance decisions and reducing engineering costs.