The use of real-world motorsport telemetry data is limited by its relative rarity, commercial confidentiality and high cost, not to mention a lack of quantity when needed for comprehensive training of aero-stability prediction systems. A Denoising Diffusion Probabilistic Model (DDPM) framework which learns a joint statistical and temporal relationship between multi-channel vehicle sensor data and produces high fidelity training sequences indistinguishable from real telemetry from a race. It features a U-Net denoising backbone with cross channel temporal attention, a linear noise schedule over T = 1000 timesteps and a physics-informed Stability Score (?) consistency loss that allows to connect the generative process to the physics of aerodynamics. The results of extensive evaluation experiments on simulated Formula-like telemetry channels, covering channels speed, acceleration, suspension movement, wings load, tyre temperature, and aerodynamic pressure, show that the proposed diffusion model outperforms GAN, VAE, and LSTM synthesis baseline models on all three metrics: 96% downstream classifier accuracy, realism score of 0.94, and mean squared error between real and synthesized channels of 0.03. The framework is intended to be an underlying basis for aero-stability artificial intelligence training pipelines that need to work on scarce data.
Vehicles with modern high performance run with complex multi-sensor data-streams under extreme aerod...
Being a collaborative, iterative process, scientific discovery requires synthesis of large amounts o...
The ground-effect Formula 1 modern cars including today's are often aerodynamically unstable which a...
Current motor sports environments can generate vast amounts of telemetry data from cameras, inertial...
