Abstract: A novel intelligent method for identifying subgrade distress was proposed to address the challenges associated with the concealed nature of subgrade defects and the scarcity of available data. The proposed framework consisted of a prediction module based on an attention-enhanced long short-term memory network(iTransformer-LSTM)and a regression module based on an elastic neural network. Mechanical response data were generated through dynamic loading tests on pavement model specimens, which effectively mitigated the problem of inadequate distress data. These data were subsequently used to predict the mechanical performance of the subgrade structure, and the distress types were identified accordingly. To verify the effectiveness of the proposed method, a scaled subgrade model was established using the material proportions of a real subgrade structure. Defective specimens representing looseness, differential settlement, cracks, and pipeline leakage were designed to reproduce realistic subgrade distress conditions. Strain-gauge earth pressure cells were embedded at different layers and positions within the model to monitor mechanical responses under dynamic loading. The measured earth pressure data were then collected, analyzed, and validated. The results demonstrated that the proposed model could accurately predict the distribution and evolution of earth pressure in the subgrade structure based on load and displacement data, thereby enabling rapid and accurate identification of distress types. The proposed method offered a new perspective for the nondestructive detection of subgrade distress.

Key words: subgrade disease detection, attention mechanism, iTransformer-LSTM, resilient neural network, model testing