Load Identification of High-Speed Train Crossbeams Using Neural Network Method: Simulated and Experimental Studies

The crossbeam is a key load-bearing structure for high-speed trains which may be damaged with long-term load. Due to the large load area of the crossbeam and the harsh service environment, it is difficult to measure the load directly. To solve this problem, a load identification method based on a neural network is proposed. First, the load area is discretized and the relationship model between load and strain is derived. Second, an extreme learning machine (ELM) improved by the sparrow search algorithm (SSA) load identification model (IELM) is derived, and strain and load are used as the input and output of the model, respectively. The method is evaluated by mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and
$R$
-square (
${R^{2}}$
), and the performance of the method is verified by simulations and experiments with the experimental MAE of 0.047 kN, RMSE of 0.086, MAPE of below 0.01, and
${R^{2}}$
of 0.995. In addition, this article compares and analyzes the proposed method with Moore–Penrose (MP) method, Tikhonov regularization with L-curve criterion method (TL), back propagation (BP) neural network method, and ELM under different noise levels. The results show that the proposed method simultaneously identifies the position and amplitude of multiple loads of high-speed train crossbeams with higher accuracy and robustness, which provides a reference for load identification of high-speed train crossbeams.