Digital Future of Product Development and Validation- The Role of Experiments & Modelling Challenges

Digital Future of Product Development and Validation- The Role of Experiments & Modelling Challenges Sam Akehurst Professor of Automotive Powertrain Systems, University of Bath

Overview Vision towards digital development Powertrain modelling challenges Future vision The role of experiments in this digital future Areas which Bath are currently working in Specific examples of activity around vehicle test that will contribute to this Conclusions

The Vision auto optimised design, verification, manufacture Business & market drivers generally focused on reducing time & cost to market with ever increasing product complexity Increased regulatory complexity will require a shift to virtual product and process certification/ homologation 70% virtual validation target for 2025

Vee process needs to move with the times Define vehicle attributes Vehicle experiment Validate vehicle Define propulsion system attributes Powertrain experiment Validate powertrain Define subsystem attributes Validate Subsystem experiment Component design & manufacture 5 Current functional decomposition typically does not consider the flexibility offered by new architectures Or Consider cross cutting technologies i.e. Thermal management of each powertrain component We need to revisit the cascading of attributes in the light of this new uncertainty We need better tools (digital and experimental) to validate powertrain system early in the process

Virtual Engineering- Powertrain Modelling Challenges Data Driven Empirical Full Physics Model Predictiveness Do we understand and have capability to model the Physics? Combustion Chemistry; Pulsating Flow in turbines, Surge, Battery degradation Many x Faster Real-time R T Many x Slower Real-time Model Execution Time Need to understand value of results and cost in time vs. accuracy Real time execution critical to HiL/SiL- control development Empirical Data is Validation Time averaged and Spatially Sparse Look up tables Neural Networks Response Surface Models???Validation Effort??? Model Fidelity Transient Thermal Models Considerable Effort Require High Spatial and Temporal resolu 3D CFD Combustion Chemical Kinetics Different fidelity solutions are required for different powertrain components and modelling solutions Fidelity depends on source of component in house developed or Tier1 supplied? 6

Future Vision: Model Creation- Fidelity Cascading- engine example New Automated Process Required Here High order models and intelligent testing 7 Parameterized low order models HW/Control optimization in system simulation/hil

Future Vision- Powertrain Development Realism Vehicle Test Correlation Rolling Road Correlation Correlation Advanced Engine Test Basic Engine Test Powertrain Simulation Cost & Complexity 8

End goal - AI led powertrain design 9 Effective selection of powertrain architecture is critical but largely left to custom and practice guided by expert knowledge Formal optimisation is needed at an early stage Places great emphasis on modelling environment More work needed on architecture optimisation with sizing and through life costing as an integrated activity Requires good enough models and good enough control optimisation to allow fair comparisons between topologies Cost Performance Emissions CO2

Vision for Powertrain Optimisation 10

Future Vision: VCHV http://vchv.uk/ Virtually Connected Hybrid Vehicle Newcastle Electric Drive Nottingham Power Electronics Model Coordination in the Cloud Loughborough Vehicle Controller Warwick Energy Storage DE&T Project Coordination Bath Combustion Engine System 11 UCL Fuel Cells

The role of experiments in this digital future Fundamental Research studies Understanding new phenomena, characterising and adding physics to modelling tools Validation data for improved modelling tools component and system level Verification and type approval Future testing will be characterised by Fewer experiments Much higher value per test Much better use of data and insights Higher volumes of more complex data 12

Near term motivation Increased powertrain diversification- multiple power sources Micro to Full hybrid RDE is the most pressing example of the need to improve experimental and analytical tools We need improved component level and system level simulation tools The validation data sets required to develop these tools are not available New experimental facilities, testing methods and model validation techniques are needed This will allow more effective system level optimisation in simulation Candidate systems will then need effective experimental validation, which is a huge challenge 13

Vehicle Characterisation for RDE The NeedValidation over all possible driving situations Analytical cycle generation Emissions Prediction System test Simulated real world drives 14 Parameterise simulation

Example of empirical approach to engine emissions modellingdynamic DoE Hot/Dynamic modelling Dynamic-Hot engine model Temperature scaling factor 3000 2000 1000 2000 Target Predicted 1500 1000 500 0 0 200 1.5 600 Time (s) 800 R2 1000 0.5 20 40 s f T 60 80 o 100 NOx Scaling 1 1 Model Validation 1 0.5 0.5 y f x, T 0 0 600 1200 0 Time (s) 0 600 Time (s) 3000 2000 1000 1200 1000 Measured Predicted 500 0 15 400 General temperature dependant model Temperature scaling function Pedal Basedmodelling input 0 Speed (rpm) Combine for general dynamic/thermal model NOx Scaling Data Pre-processing Hot/dynamic model validation NOx (ppm) Cold start data acquisition y f x Speed (rpm) Varying frequency sine waves (Chirp signals) Hot start NEDC Cold start NEDC Data Pre-processing Cold start test design Dynamic training sequence Warm-up behaviour a challenge Hot engine data acquisition NOx (ppm) Volterra series for mechanical dynamics Scaling factor for thermal effects Hot engine test plan NOx Scaling Dynamic modelling approach Problem Definition Dynamic inputs Temperature input Emissions output 0 200 400 600 Time (s) 800 1000 1200 1200

Advanced Chassis/Powertrain dynamometer 17

Requirements for the Chassis/Powertrain dynamometer All of the precision, control of an engine dyno but with the boundary conditions of an RDE chassis dyno Consider altitude simulation, state of charge, after-treatment state + parasitic loads Improved robot driver control and integration Improved instrumentation to give high bandwidth, system wide data Engine torque, Axle torque, Fuel mass flow Fast emissions, Electrical system Control over engine actuators, speed, load Necessary to allow separation of physical responses from controller imposed behaviours. We need to model the former and include the latter in the downstream optimisation Full integration with the optimisation suite (and accessible by calibration engineers) Implies a relatively mature powertrain and mule vehicle is available 18

Understanding the real world road test Total On-road info: EU6 Diesel (SCR) Repeated routes throughout testing period 4 drivers CAN Bus monitoring Replicate on CD Trip Duration (s) Trip Distance (km) Trip Average Vehicle Speed (km/h) Mean St. dev. Min Max 3680 44 3460 45 60 0.1 21150 208 39 17 2 88 10 6-2 29 Trip Average Ambient Temperature ( C) 20 Driver No. of Trips All Data 314 1 80 2 115 3 70 4 49 Total Distance (km) 13700 3330 3680 4430 2260 Average Speed (km/h) 39 44 30 47 37 Average trip distance (km) 44 42 32 63 47

WLTC vs. NEDC vs. On Road NEDC has distinct hot spots Steady state On-road driving covers much broader range Hot spots exist in narrow speed ranges (cruising) but cover wide torque regions (varying road load for same speed) 21 WLTC Has broader coverage - Transients

On-Road Torque 22 Driver to Driver Variation On-road driving covers much broader range Hot spots exist in narrow speed ranges (cruising) but cover wide torque regions (varying road load for same speed)

Temperatures around after-treatment are critical Coolant temperature Oil temperature Pre-cat exhaust temperature Post-cat exhaust temperature Post-DPF exhaust temperature Blue = WLTC Black = NEDC Traces that start high are hot restarts 23

More cycles alone will not be enough Operating boundaries of RDE are very wide compared with NEDC, real word is wider still Physical processes across powertrain are highly nonlinear Interactions between powertrain subsystems not well represented in today s system level models Modern control strategies contain many discontinuities Switching behaviour (with temperature etc) Map based (gradient is discontinuous and map data subject to calibration process errors) Cycle based approach does not define driver behaviour 24 Together these factors mean that a cycle based approach can never deliver exhaustive validation

Possible RDE optimisation workflow Hardware selection Advanced steady state test Optimise for full map steady state compliance Develop models robust to RDE boundary conditions Advanced powertrain and subsystem test Generate full powertrain dynamic system models Dynamic simulation of real manoeuvres Optimisation of calibration 25 Digital Dynamic characterisation on CD Measured Predicted 200 400 600 Time (s) 800 Validation testing on CD over preset cycles Validation testing on road with PEMS Experimental

Steady state test enhancements Adding in the requirement to test at many more conditions to reflect RDE conditions More capable cells in terms of environment, control and instrumentation Workload unmanageable without improvements in test design and execution 26 Incorporation of prior knowledge to speed up limit search Iterative on-line DoE to minimise data requirement Sweep mapping to yield data more rapidly Bayesian techniques to incorporate prior knowledge into response models Dynamic DoE? Bath have demonstrated techniques to do each of these steps, needs exploitation

Iterative online DoE Process Start with a simple DoE design Select next points to test based on points of least confidence Depends upon close real time integration between cell and DoE tool Recalculates models of mean and variance on the fly Also opportunity to use a Bayesian approach to reduce convergence time and improve reuse of simulation or historical data Limit search can be improved in a similar way Improved interpolation techniques such as natural neighbours can improve fidelity of models 27

Simulation requirements for optimisation Greater insight into system behaviour Multi-physics, dynamic models, still with data driven elements for the most non-linear features (combustion, emissions) until techniques improve Well defined modelling framework allowing model interaction Realistic driver and environment models Robust and comprehensive validation data sets! this is where most of our future experimental effort will be placed 28

Wider challenges Perhaps the biggest challenge is the way large companies traditionally work Design, simulation, manufacturing, test ops., calibration all need to be joined up Re-use and improve the initial models throughout the process Use models to guide experiments and the data to improve the modelsclose the loop Better capture and use of in-service data Most engine/powertrain dynos and chassis dynos are not flexible or precise enough today Neither are their operating practices 29

Conclusions Long term goal is digital design and validation In the near term, RDE will require systematic engine and powertrain optimisation in vehicle system context Behaviour dominated by steady state capability Boundary conditions and interactions critical Longer term, hybridisation challenges and complexity of thermal management Calibration and Signoff against CD cycles unlikely to be a robust process on its own Better use of software tools is essential Use the CD/Powertrain Dyno to generate a rich dataset Validate advanced models Optimise in software Signoff on random or worst case cycles with more confidence Significant implications for test design/operation Precision, Environment Access to engine data and actuators 30

Any Questions? Sam Akehurst Professor of Automotive Powertrain Systems, Institute of Advanced Automotive Propulsion Systems University of Bath Bath BA2 7AY S.Akehurst@bath.ac.uk http://www.iaaps.co.uk Thanks to my colleagues who have contributed to this presentation through their research. 31