On machine learning, system identification and internet-distributed validation of powertrains

  • Adria Ametller Picart

Student thesis: Doctoral ThesisPhD


Amongst the myriad of potential hybrid powertrain architectures, selecting the right one for a given application is a daunting task. Whenever available, computer models greatly assist in the task. However, some elements, such as pollutant emissions, are difficult to model, leaving no other option than to test, for which at some point a real powertrain will be needed. Validating plausible options before assembling the entire powertrain has the potential of speeding up the development of vehicles. Doing so without having to ship the components around the world, even more.

This work undertakes the task of designing a system to link test rigs over long distances in order to virtually couple vehicle components whilst avoiding physical contact. In the past, methods have been attempted with and without using mathematical models of the components to couple. In both cases the methods show reasonable accuracy only when the systems to couple present slow dynamics in relation to the communications delay. In addition, these methods seem to overlook the implications of operating a distributed system without a common time frame with synchronized clocks, as no method explicitly accounting for setpoint synchronisation has been found. Therefore, the problem of remotely coupling highly dynamic components remains still unsolved.

In order to overcome the inherent latency arising from long-range communication, the proposed design combines the two following features in a novel arrangement: The exploitation of synchronised clocks to introduce setpoint commands simultaneously, and the use of models (observers) of the components being coupled, generated through their own operational data.

Despite the appeal of observer-free coupling techniques, these are deemed limited in their ability to predict future behaviour under all circumstances, since these are generally based on some sort of static predicting rule/filter based on immediate past behaviour. The situation is analogous to that of driving a car while watching the rear-view mirror. It works well when the road ahead is straight, but not so well when curves lie in front. Hence, the observer method route is preferred. Nevertheless, the use of models clashes against the essence of the application – if good models were available, why not just simulate the coupling on a computer? This dilemma is sought to be minimised by using data-driven models requiring no prior plant knowledge. These models are created using the LOLIMOT algorithm.

The designed coupling architecture is tested against two simulated physical systems. The first one, a simple deterministic system consisting of three rotating inertias coupled by means of spring-dampers, in which over a 70% error reduction is obtained when compared to direct transmission of the signals. The second one, consists of an internal combustion engine coupled to an electric motor/generator, typical of a hybrid vehicle powertrain configuration. Despite improvement over the duration the coupling can be kept working compared to direct transmission of the signals – 60 seconds against 20 seconds –, the fidelity of the virtual coupling remains far from faithful to the physical behaviour. The reason lies in the quality of the observers obtained through the LOLIMOT algorithm, especially that for the engine. Different methods of data collection are devised to improve these models, finding that data stemming from the original physical coupling results in better models. However, having to do so goes against the nature of the application. As a result, it is concluded that the LOLIMOT algorithm is inadequate to model an engine as a single unit for the objective of the application. Nevertheless, the devised coupling architecture may still prove useful in the event of obtaining more accurate models. Although from the point of view of the application, having to employ physics-based models is not as desirable as pure data-driven models, the blending of the two, in the so-called hybrid models, may be the most promising route to success.
Date of Award12 Dec 2022
Original languageEnglish
Awarding Institution
  • University of Bath
SponsorsDigital Engineering and Test Centre
SupervisorChris Brace (Supervisor), Richard Burke (Supervisor) & Simon Pickering (Supervisor)


  • Hardware-in-the-Loop
  • Internet-Distributed HiL
  • X-in-the-Loop
  • Latency mitigation
  • Powertrain validation
  • Machine learning
  • System identification
  • Modelling
  • hybrid electric vehicle (HEV)
  • Residual analysis

Cite this