Construction of risk spectrum for vehicle stability state based on data-driven

• DBSCAN algorithm clusters vehicle states into four types: stable, tend towards stable, tend towards unstable, and unstable. • A 1D-CNN is trained on the clustered data to construct a vehicle stability risk spectrum based on classification probability. • Various methods verify that the risk spectrum is useful in quantifying the probability of vehicle stability and identifying the stable state. • This work intends to inform active safety control system design and broaden current dynamic control safety limits. This study develops a data-driven vehicle stability risk spectrum to quantify safety risks and accurately identify stability states. A dataset comprising multi-dimensional vehicle state attributed is first generated using driving simulators. Subsequently, the density-based spatial clustering of application with noise (DBSCAN) algorithm is employed to classify vehicle states into four categories: “stable”, “tend towards stable”, “tend towards unstable” and “unstable”. These clustered data are then used to train a one-dimensional convolutional neural network (1D-CNN) to construct the vehicle stability risk spectrum based on classification probabilities. Finally, using the Lyapunov index method, nonlinear system phase plane method, Carsim/Simulink platform, and hardware-in-the-loop (HIL) testing, typical stability states are tested and verified. The results show that the vehicle stability risk spectrum constructed can effectively predict the state of vehicles and calculate the safety and stability boundary. This paper aims to timely and accurately identify dangerous states of vehicles, provide a reference for active safety control system design, and expand dynamic control safety boundaries.