Abstract
A digital twin is a digital representation of a planned or real physical system, product, or process that functions as its practically identical digital counterpart for tasks such as testing, integration, monitoring, and maintenance. Creating digital twins allows the ‘digital system’ or ‘digital product’ to be tested faster-than-real-time improving the overall efficiency and reducing programme’s timescales.
The HORIBA Intelligent Lab virtual engineering toolset has been used produce Empirical Digital Twins (EDT) of several contemporary Internal Combustion Engine (ICE) propulsion systems. Digital twinning in its current format can supplement real-world testing methods for the development, calibration, optimisation, and certification of powertrains and vehicles. Given the move to ever more stringent pollutant criteria over wider test conditions, which is expected to put additional strain on Original Equipment Manufacturers (OEMs), the HORIBA EDT approach is expected to increase efficiency and reduce time when taking a vehicle or powertrain to market.
The flexibility of the HORIBA EDT approach also means it is highly suitable for other automotive propulsion systems including fuel cells, batteries, and invertors in addition to non-automotive systems such as water quality and electrical systems and components.
Dr. Phil Roberts
Technical Specialist,
Propulsion Research and Development, HORIBA MIRA Ltd.
phil.roberts(at)horiba-mira.com
Dr. Luke Bates
Senior Engineer,
Data Driven Innovation,
HORIBA MIRA Ltd.
luke.bates(at)horiba-mira.com
Kunio Tabata
Manager,
HORIBA, Ltd.
kunio.tabata(at)horiba.com
Steve Whelan
Global Application Centre Lead - Emissions,
Intelligent Lab,
HORIBA MIRA Ltd.
steve.whelan(at)horiba-mira.com
► Click Here for original paper (PDF file)
Introduction
A paradigm shift in the development of vehicles and powertrains will occur due to the introduction of more stringent emissions limits (EU7 in Europe is a prime example). At the time of writing, vehicles sold in Europe must still go through laboratory testing using the Worldwide Light-Duty Test Procedure (WLTP) and on-road Real Driving Emissions (RDE) regulations to comply with EU6 emissions regulations. The WLTP is undertaken under controlled laboratory conditions that are often favourable for engine performance and emissions abatement technologies, with on-road RDE testing being more complex and subject to more variable conditions such as driving criteria and route characteristics (urban, rural, and motorway driving) that are still well defined by the boundary conditions of EU6. Although the RDE test conditions are often challenging to meet, the clearly defined boundaries mean that the OEMs can develop their powertrains and aftertreatment abatement technologies to operate effectively within explicit on-road conditions.
Achieving sufficient testing coverage to guarantee emissions compliance for more stringent emissions limits over wider operational boundaries using physical testing alone will be challenging, if not practically impossible for all OEMs. It is likely therefore that simulation toolsets that can be used to predict performance and emissions attributes for difficult-to-achieve and edge-case road drives will be paramount in trying to achieve compliance with future emission standards. In this respect, the authors expect OEMs to physically drive their ‘worst case’ road tests and then use simulation and prediction to supplement physical road testing as shown in Figure 1. Here, for a single powertrain and vehicle derivative adopting a single RDE route incorporating 3 different traffic conditions, 3 different driving styles, 3 different battery States of Charge (SOC), and 4 different environmental conditions, the number of physical tests possible is more than 100. This number of physical tests is exceptionally difficult to achieve to a satisfactory standard, therefore supplementary data from simulations will be vital.
To support OEMs in achieving compliance with future emissions regulations, a suite of tools has been developed within the HORIBA Intelligent Lab working group at HORIBA MIRA that utilises the power of digital twinning – constructing an accurate, digital representation of a physical unit using empirically derived models. The HORIBA EDT toolset utilises in-house developed transient design, modelling, prediction, and optimisation modules similar in context to a transient Design of Experiments (DOE) methodology. These standalone modules are then used to: 1.) exercise powertrains and/or vehicles in an explicit, but non-typical manner to generate empirical training data, 2.) generate empirical performance and emissions models using the transient training data, 3.) predict performance and emissions data for real-world or synthetic road drives and/or manuovers, and 4.) identify hotspots – unfavourable powertrain attributes using simulation or real-world generated driving or vehicle operating conditions that produce excessive emissions or high fuel or energy consumption.
The approach detailed above has been used to create EDTs of several powertrains tested on engine, powertrain, and chassis dynamometers; Figure 2 and Figure 3 are examples of techniques used to create an EDT of two different powertrains. The EDTs are then coupled with either 1.) simulated real-world driving cycles derived from the IPG CarMaker virtual vehicle simulation toolset (or similar programme) and/or 2.) real-world driving or machine handling cycles; the latter with respect to non-road mobile machinery. The HORIBA EDT approach is an evolution of some of the quasi-dynamic testing methodologies outlined in [1-6] yet expands on these approaches using novel modelling techniques combined with simulation to create a more immersive and greater depth of field and solutions.
Figure 1 An example for the need for combined physical testing and simulation.
Figure 2 PHEV powertrain tested at HORIBA MIRA in the UK.
Figure 3 Vehicle and corresponding diesel engine tested at an OEM in the UK.
Operational Principle – Overview
The HORIBA Intelligent Lab transient EDT toolset was used to generate the training cycles to excite both a PHEV powertrain and commercial vehicle powertrain across their entire operating ranges. This meant that the transient behaviour of both powertrains would be represented in the empirical performance and emissions models that were created using the recorded training data. The excitation training cycle was approximately 1.5 hours in length for both cases. The accelerator pedal, brake pedal, and the rotational wheel speed for the PHEV powertrain were used as inputs to its training cycle with accelerator pedal and vehicle velocity used as inputs for the commercial vehicle.
The latter was tested using a patented technology from the Intelligent Lab working group at HORIBA MIRA called Torque Matching™ whereby a chassis dynamometer can be operated in ‘speed mode’ in the same way an engine dynamometer can be operated. This means that the chassis dynamometer controls the vehicle speed (no need for coastdown coefficients or other factors associated with ‘driving’ on a chassis dynamometer) with a robot driver actuating the accelerator pedal.
The training data captured for both powertrains was used to create transient empirical models (EDT) of several performance and emissions attributes. These models were then validated for accuracy before they were used to predict performance and emissions attributes for representative real-world driving scenarios that were generated using the IPG CarMaker virtual vehicle simulation toolset. Some further information about the experimental design, training data, model generation, model validation, and model prediction are discussed in the following subsections, with an outline of the empirical digital twinning DOE approach presented in Figure 4 and Figure 5.
Figure 4 Stages of the HORIBA EDT approach (1); experiment design, generation of training data, modelling, and model validation.
Figure 5 Stages of the EDT approach (2); establishing real or virtual scenarios to couple with empirical models, prediction of responses, and ‘hotspot’ determination.
Operational Principle – EDT Excitation Signal Design
The described EDT method relies on machine learning and Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM). This type of learning typically requires large volumes of data to train a reliable and accurate model, since the data used is typically relatively information sparse, as with real road test data or drive cycle data. To reduce the volume of data required, and so the physical test burden, a dynamic excitation signal specific to the system under test is designed. The design attempts to cover the whole operating space of the system while capturing dynamic system behaviour and increasing the density of information in the data produced.
The excitation signal consists of a series of target points within the constraints of system operation, similar to a steady-state DOE. These points are traversed with only brief pauses (if any) at the specified conditions, and points may be revisited multiple times in varying sequences. Unlike a steady-state DOE the sequence of points is important. A dynamic system may produce different behaviours at the same condition depending on the approach trajectory to that condition i.e., the ‘history’ leading up to the measurement value is important and thus accounted for in the modelling approach. By covering the entire operating space of the system, the produced data can be applied to model any given scenario that the system may encounter.
Operational Principle – EDT Model Creation, Validation, and Prediction
RNN methodology is chosen because of its proven ability to deal with dynamic systems with multiple timescales. This is complemented by LSTM, improving this ability for extended timescales. The model accounts for the dynamics of the system by using sequences as inputs; the point predicted is at the end of this sequence. This provides the capacity for the same end-point conditions to generate a different response depending on the sequence of points preceding the prediction. LSTM modules additionally store some values and can reintroduce these values to the network, even if they are not generated from the current sequence. These values are forgotten by the modules if a learned condition is met.
The learning algorithm is provided with training data to train the model(s). This data includes both the input variables that will be required to utilise the trained model and the variables to be predicted, which are recorded during the designed test. Data is applied to the algorithm as a sequence, with the sequence being a moving window that advances one data point per iteration of training.
Training an RNN is potentially time-consuming. However, by leveraging both information-dense data and a General-Purpose Graphics Processing Unit (GPGPU), this is reduced to several minutes of training with prediction taking a matter of seconds for a 90-minute drive scenario with several variables provided at 10Hz. Also note that all models generated are subject to validation i.e., comparing predicted responses with real measured data. An example of the model quality that can be produced with the HORIBA EDT toolset is shown in Figure 6 with the difference between measured cycle results and predicted cycle results displayed for several performance and emission attributes.
Real-World Scenario Generation
For the majority of the EDTs created thus far, the IPG CarMaker virtual vehicle simulation software was used to create realistic real-world driving scenarios. From the work outlined in [7], the virtual driver was programmed to drive in an explicit manner across several virtual RDE cycles that were programmed within IPG CarMaker. These real-world driving scenarios can be resolved orders of magnitude faster than real-time allowing for several hundred/thousand scenarios to be resolved for a single vehicle in a very short timescale. Some of the vehicle and/or powertrain data from these simulations is subsequently coupled with the EDT of the corresponding vehicle/powertrain; thereby allowing the performance and emissions to be predicted for a huge number of cycles without the requirement for significant on-road testing. An example of the types of scenarios that can be generated using IPG CarMaker are shown in Figure 7.
Figure 6 Measured vs. predicted deltas for performance and emissions attributes for a PHEV SUV tested on a powertrain dynamometer.
Figure 7 IPG CarMaker scenario definition for the light-duty PHEV SUV.
After predictions, the data can be interrogated for trends in engine/powertrain/vehicle performance with respect to such characteristics as driving style, traffic density, and route characteristics. In addition, if the EDT models incorporate the effects of altitude and cold/hot ambient running conditions, performance and emissions can be predicted for virtual routes ‘driven’ above sea-level and at extremes of nominal ambient temperatures. An example of this is the work outlined in [8] which utilised the HORIBA MEDAS system to incorporate the effects of altitude and temperature on engine performance and emissions within the EDT.
Hotspot Identification
The forthcoming EU7 emissions regulations scheduled for introduction in the middle of this decade will introduce a not-to-exceed mass limit budget in addition to the established cycle limit; the former being introduced specifically to negate the beneficial effects of the prolonged warm/hot running during road testing which inherently ‘dampens’ the overall effect of cold start emissions on the cycle result.
An example of how the EDT approach can be used to identify hotspots – unfavourable engine or vehicle operating conditions or road tests – is shown in Figure 8. In this figure are the predicted cumulative tailpipe NOx emissions for 4 RDE cycles adopting the same virtual route and traffic density. Battery SOC for the SUV PHEV model was zero in each case. The black lines represent predictions made at sea level and 35°C intake temperature for gentle (solid) and dynamic (dashed) driving with the red lines indicating predictions made at 1800m, and 35°C intake temperature for the same gentle (solid) and dynamic (dashed) driving conditions.
The black box indicates the proposed not-to-exceed EU7 tailpipe NOx mass limit of 600mg within the first 10km of the test. All predictions are made with training data captured with a ‘hot’ engine.
The cumulative traces for driving at sea level (black lines) pass through the vertical plane of the mass limit box (~400mg within the first 10km) and go on to produce less than 60mg/km tailpipe NOx – the proposed EU7 limit for tailpipe NOx for both gasoline and diesel. Conversely, the predictions made for driving at 1800m (red lines) depart out of the horizontal plane of the mass limit box generating 600mg NOx within 6-7km. Interestingly, although driving at high altitudes results in a test failure regarding the not-to-exceed mass limit, these cycles produced less than 60mg/km which would according to EU6 regulations denote a successful test.
The data shown in Figure 8 is an example of how the HORIBA EDT methodology can identify problematic operating conditions for a powertrain before signoff is complete. Note also that these predictions were made using training data that was recorded with a ‘hot’ engine. The results presented here are therefore likely to change significantly when cold start is accounted for during both sea-level and high-altitude driving, with often potentially favourable operating conditions resulting in tailpipe NOx greater than 600mg within the first 10km. Therefore, it is the view of the authors that cold starting conditions, whilst problematic from an EU6 perspective but mitigated somewhat by the prolonged ‘hot’ driving and adherence to a total cycle result, will be the dominant contributing factor to meeting the EU7 emissions regulations due to the addition of the 10km mass limit.
Figure 8 Predicted tailpipe NOx emissions for the PHEV SUV at sea-level and 1800m.
Conclusions
The introduction of more stringent emissions regulations will require OEMs to undertake increased simulation activities to ensure their products are compliant across broader ranges of operational conditions. HORIBA’s Intelligent Lab transient EDT methodology has been effectively used to predict performance and emissions for several powertrains, engines, and vehicles. By adopting an EDT approach, OEMs can ensure their products are robust and will meet emissions regulations under ‘worst-case’ or ‘edge-case’ scenarios that may be difficult, if not practically impossible to achieve during real-world road testing. Furthermore, with powertrain sharing commonplace amongst OEMs, if the powertrain can be exercised via an explicit excitation signal across an operating range that suits all vehicles it is used in, then the performance and emissions models created can be coupled with data extracted from scenarios that utilise several different vehicle models. Some preliminary results indicate that adopting an EDT approach for a single vehicle can reduce verification and validation signoff timescales by at least 70% with this increasing as multiple vehicles adopt a common powertrain.
The EDT approach outlined here has been deployed to vehicles already in production with well-established powertrains. However, this approach can readily be used for immature powertrains to optimise hardware and calibrations before sign-off is required. For example, an EDT solution can be implemented during air handling or aftertreatment selection programmes in order to reduce the number of test prototypes required.
Whilst the current work has focused on liquid-fuelled conventional and electrified powertrains, the EDT approach presented here is also compatible with fully electric vehicles (fuel cell or battery) to understand how attributes such as energy consumption and range can be affected by different driving dynamics, environmental conditions, or payloads. This is particularly useful given the often-laborious procedure of quantifying energy consumption and range using traditional chassis dynamometer techniques. Furthermore, the flexibility of an EDT approach means that other non-automotive systems can also benefit from this approach as the toolset development is cross-functional across many engineering and scientific areas.
Acknowledgements
The authors wish to thank the global HORIBA teams for their contributions to the work published in this paper.
References
[1]Gutjahr, T., Kruse, T., and Huber, T., “Advanced Modelling and Optimization for Virtual Calibration of Internal Combustion Engines,” NDIA Ground Vehicles Systems Engineering and Technology Symposium, 2017.
[2]Shishido, T., He, J., Kaihatsu, M., Haukap, C., Dreher, T., and Hegmann, M., “Dynamic Modelling for Gasoline Direct Injection Engines,” Keihin Technical Review, Volume 6, 2017.
[3]Le Corre, S.D., Mason, B., Steffen, T., Winward, E., Yang, Z., Childs, T., Cary, M., and Lygoe, R., “Application of Multi-Objective Optimization Techniques for Improved Emission and Fuel Economy over Transient Manoeuvers,” SAE Technical Paper 2019-01-1177, 2019, DOI:10.4271/2019-01-1177.
[4]Tulpule, P, Rezaeian, A., Karumanchi, A., and Midlam-Mohler, S., “Model Based Design (MBD) and Hardware In The Loop (HIL) Validation: Curriculum Development,” American Control Conference, 2017, DOI:10.23919/ACC.2017.7963788.
[5]Xiaohang, F., Papaioannou, N., Leach, F., and Davy, M.H., “On the application of artificial neural networks for the prediction of NOx emissions from a high-speed direct injection diesel engine,” International Journal of Engine Research, Volume 22, Issue 6, Pages 1808-1824, 2021, DOI:10.1177/1468087420929768.
[6]Katreddi, S.; Thiruvengadam, A., “Trip based modelling of fuel consumption in modern heavy-duty vehicles using artificial intelligence,”. Energies 2021, 14, 85–92, DOI: 10.3390/EN14248592.
[7]Roberts, P.J., Mason, A., Tabata, K., and Whelan, S., “RDE Plus – A Road to Rig Development Methodology for Whole Vehicle RDE Compliance: Engine-in-the-Loop and Virtual Tools,” SAE Technical Paper 2020-01-2183, 2020, DOI:10.4271/2020-01-2183.
[8]Roberts, P.J., Bates, L., Tabata, K., Headley, A., and Whelan, S., “RDE Plus – Rapid Characterisation of Vehicle and Powertrain Performance and Emissions using Dynamic Design of Experiments, Digital Twin and Virtual Driving Methodologies,” SAE Technical Paper 2022-01-0580, DOI: 10.4271/2022-01-0580.
About HORIBA MIRA
HORIBA MIRA is a global provider of automotive engineering, test and development consultancy. We work in close collaboration with vehicle manufacturers and suppliers around the world, providing world-class support for automotive product development.
From technology development through to full-vehicle programmes and certification, we have the breadth, capability and flexibility to support an international customer base, with our engineers utilising the latest test facilities and simulation tools to improve lives by making journeys safer, cleaner and smarter.
We’ve been developing some of the world’s most iconic vehicles for nearly 80 years, working to achieve our vision – that every journey in the world will be positively influenced by us. The unique combination of our advanced facilities, engineering expertise and prime location of MIRA Technology Park in the heart of the UK automotive industry, ensures our customers can develop and validate their technology, or vehicle, in one place.
Enquirie : https://www.horiba-mira.com/enquiries/
e-Readout Editorial Office:
Research & Development Division, HORIBA, Ltd.
2, Miyanohigashi-cho, Kisshoin, Minami-ku, Kyoto 601-8510, Japan
