Control Sequences Generation for Testing Vehicle Extreme Operating Conditions Based on Latent Feature Space Sampling

Extreme operating conditions refer to the critical dynamic state during vehicle operation. The lack of experimental data under critical conditions is one of the fundamental problems in the study. To solve the problem, we design an LSTM-VAE based generating model to generate rational control sequences that can push vehicles toward extreme operating conditions and used simulation tests to analyze them. Specifically, we train the Encoder to study the basic driving logic of the control sequences collected during free-drive tests by human drivers, forming a low-dimension latent feature space. Then, we sample from specified regions in the latent feature space and use the Decoder to generate new control sequences. Finally, we use the sequences as the control input of the 27-DoF high-precision vehicle dynamic simulation platform and analyze the variations of simulated vehicle dynamics. We conduct different experiments and validate the method from different aspects. Results reveal that by sampling from specific regions of the latent feature space, we get a higher chance to generate desired control sequences for extreme operating conditions.