Abstract: Real-time damage prediction and life assessment of vehicle structures require dynamic load information of each component. However, the dynamic load identification method based on the test method and the traditional time-frequency domain has certain limitations in terms of identification accuracy and stability. A vehicle dynamic load prediction method is proposed. Through a small amount of input signals, the dynamic load of the main failure risk components such as rear axle, frame and leaf spring is predicted, which lays the foundation for the realization of low-cost dynamic monitoring of vehicle damage and intelligent life management. Firstly, the actual collected signal is subjected to noise reduction processing by CEEMDAN-PE method, and the denoised wheel center acceleration signal and displacement signal are used as input, and hyperparameters are adjusted based on training data and independent test data to establish the optimal architecture NARX neural network model. The strain load of each component is predicted in real time, and the component damage is calculated in real time, and the prediction results are evaluated from the time domain, frequency domain, damage domain and other aspects. On this basis, combining the input and output data under different road conditions to build the NARX-S single road condition model and the NARX-M multi-road condition unified model, the comparison of multiple models under different road conditions shows that the NARX-S single road condition model has higher prediction accuracy, while the NARX-M prediction accuracy is relatively lower, but has better general adaptability.

Key words: dynamic load, neural network, noise reduction processing, load prediction, damage calculation