A Real-Time Wheel–Rail Force Identification Method via Axlebox Acceleration-Based Res-FBLM Model

Wheel–rail force is a critical physical parameter in rail transit, playing a vital role in assessing both the operational safety of vehicles and the durability of structural service. However, existing studies have shown that wheel–rail force identification requires multiple types of data input, which poses a significant challenge for engineering applications. To address this issue, this article proposes a residual frequency-enhanced bidirectional LSTM with multihead attention mechanism (Res-FBLM) model, which relies solely on triaxial signals from the axlebox as input. Subsequently, a multibody dynamics model of an urban railway vehicle is established and simulated under various conditions to generate wheel–rail force and acceleration data. The research results show that the proposed model achieves high-precision identification of wheel–rail forces, with an $R$ -squared ( $R^{2}$ ) value exceeding 98% on the test set. Additionally, the proposed method was further validated on the wheel–rail interaction test rig (WRITR). The computation time on the edge deployment device, a Raspberry Pi, is 12.4 ms/sample, which satisfies the real-time constraints imposed by high-frequency sampling. Research indicates that the proposed model achieves high accuracy and real-time performance using only axlebox acceleration as input, thereby offering technical support for the safety monitoring of railway vehicles.