Demand-Driven Data Acquisition for Large Scale Fleets.

Automakers manage vast fleets of connected vehicles and face an ever-increasing demand for their sensor readings. This demand originates from many stakeholders, each potentially requiring different sensors from different vehicles. Currently, this demand remains largely unfulfilled due to a lack of systems that can handle such diverse demands efficiently. Vehicles are usually passive participants in data acquisition, each continuously reading and transmitting the same static set of sensors. However, in a multi-tenant setup with diverse data demands, each vehicle potentially needs to provide different data instead. We present a system that performs such vehicle-specific minimization of data acquisition by mapping individual data demands to individual vehicles. We collect personal data only after prior consent and fulfill the requirements of the GDPR. Non-personal data can be collected by directly addressing individual vehicles. The system consists of a software component natively integrated with a major automaker's vehicle platform and a cloud platform brokering access to acquired data. Sensor readings are either provided via near real-time streaming or as recorded trip files that provide specific consistency guarantees. A performance evaluation with over 200,000 simulated vehicles has shown that our system can increase server capacity on-demand and process streaming data within 269 ms on average during peak load. The resulting architecture can be used by other automakers or operators of large sensor networks. Native vehicle integration is not mandatory; the architecture can also be used with retrofitted hardware such as OBD readers.

Strong dynamical heterogeneity and universal scaling in driven granular fluids.

Large-scale simulations of two-dimensional bidisperse granular fluids allow us to determine spatial correlations of slow particles via the four-point structure factor S(4)(q,t). Both cases, elastic (ϵ=1) and inelastic (ϵ<1) collisions, are studied. As the fluid approaches structural arrest, i.e., for packing fractions in the range 0.6≤ϕ≤0.805, scaling is shown to hold: S(4)(q,t)/χ(4)(t)=s(qξ(t)). Both the dynamic susceptibility χ(4)(τ(α)) and the dynamic correlation length ξ(τ(α)) evaluated at the α relaxation time τ(α) can be fitted to a power law divergence at a critical packing fraction. The measured ξ(τ(α)) widely exceeds the largest one previously observed for three-dimensional (3d) hard sphere fluids. The number of particles in a slow cluster and the correlation length are related by a robust power law, χ(4)(τ(α))≈ξ(d-p)(τ(α)), with an exponent d-p≈1.6. This scaling is remarkably independent of ϵ, even though the strength of the dynamical heterogeneity at constant volume fraction depends strongly on ϵ.

Anomalous velocity distributions in active Brownian suspensions.

Large-scale simulations and analytical theory have been combined to obtain the nonequilibrium velocity distribution, f(v), of randomly accelerated particles in suspension. The simulations are based on an event-driven algorithm, generalized to include friction. They reveal strongly anomalous but largely universal distributions, which are independent of volume fraction and collision processes, which suggests a one-particle model should capture all the essential features. We have formulated this one-particle model and solved it analytically in the limit of strong damping, where we find that f(v) decays as 1/v for multiple decades, eventually crossing over to a Gaussian decay for the largest velocities. Many particle simulations and numerical solution of the one-particle model agree for all values of the damping.

Slow dynamics and precursors of the glass transition in granular fluids.

We use event driven simulations to analyze glassy dynamics as a function of density and energy dissipation in a two-dimensional bidisperse granular fluid under stationary conditions. Clear signatures of a glass transition are identified, such as an increase of relaxation times over several orders of magnitude. As the inelasticity is increased, the glass transition is shifted to higher densities, and the precursors of the transition become less and less pronounced, in agreement with a recent mode-coupling theory. We analyze the long-time tails of the velocity autocorrelation and discuss its consequences for the nonexistence of the diffusion constant in two dimensions.

Long-time tails and cage effect in driven granular fluids.

We study the velocity autocorrelation function of a driven granular fluid in the stationary state in three dimensions. As the critical volume fraction of the glass transition in the corresponding elastic system is approached, we observe pronounced cage effects in the velocity autocorrelation function as well as a strong decrease of the diffusion constant, depending on the inelasticity. At moderate densities the velocity autocorrelation function is shown to decay algebraically in time, like t(-3/2), if momentum is conserved locally, and like t(-1), if momentum is not conserved by the driving. A simple scaling argument supports the observed long-time tails.