Physics-informed Artificial Intelligence for Cutting Brake Emissions from Electric Vehicles – PI-CUBE

Two approaches were initially pursued in parallel: on the one hand, the exploitation of existing data from industrial test benches, and on the other, an extension of the functionalities of a highly instrumented laboratory test bench (tribometer).
The use of industrial test data with learning algorithms was carried out on noise pollution (squeal), as there is currently insufficient data on the measurement of particle emissions. These data proved insufficient for a very robust prediction.
This is the challenge of laboratory tribometer testing, the main advantage of which is that it incorporates a wealth of instrumentation enabling us to identify the main mechanisms and the associated data. The first step was to define test conditions that would enable us to reproduce loads close to the real case (in terms of forces, temperature, etc.) while obtaining a satisfactory emissions response. A selection of friction materials was carried out during an initial experimental campaign, which also enabled us to familiarize ourselves with the test bench and its multimodal instrumentation. Two pairs of friction materials were selected according to 6 criteria deemed essential for the project, with distinct tribological responses and sensitivity to emissions.
Several modifications were made to the test set-up and instrumentation, making the system more reliable and increasing the number of measurements (in particular, an operando particle collection and measurement chain was added, as well as discrete surface monitoring, i.e. between each test, using two optical profilometers on the pawn and disc sides).
Various classification algorithms were used to record and classify both noise and particle emission episodes. Prediction algorithms were tested to predict emissions, and to identify influential parameters. Cross-references are made with physical interpretations to enrich predictions.
The development of physics-driven learning predictions was finally used for real-time prediction on the braking tribometer. An initial learning phase is carried out, then the predictor is used to modulate the friction brake actuator with the aim of reducing emissions. This modulation takes into account the possible combinations between the friction brake and the regenerative (electric) brake.

The use of data from industrial test benches with a large amount of data has shown that a relatively satisfactory prediction can be obtained with reliable interpolation within a defined operating range (speed, pressure, etc.). Nevertheless, these predictions are linked to a given application (type of brake), with limited robustness and no extrapolation to other braking configurations due to a lack of measured data and understanding of the associated predominant physical mechanisms.
A highly instrumented laboratory tribometer test configuration was developed, combining operando mechanical and thermal measurements with quantification of noise and particle emissions (counting and collection) and discrete monitoring of surfaces in contact (between each braking action).
In-depth analysis of the data was carried out from a physical point of view, through interpretation and modeling of phenomena identified as key to emissions: contact opening/closing, role of temperature, initial surface conditioning, etc.). These key parameters were confirmed using feature importance methods. This confrontation between physical interpretation and identification of key parameters proved highly satisfactory, shedding light on both the predominant physical mechanisms and the methods needed to predict these nuisances, both noise (squeal) and particle emissions.
Based on the classification of emissions and the identification of key parameters, learning algorithms were developed and tested over several campaigns. The result is a campaign with real-time prediction of emissions using the algorithm developed, with learning from previous campaigns. The result is an emissions predictor that can be used to adjust braking effort to limit emissions, while guaranteeing the dissipation of the energy fraction of the mechanical brake, in addition to the regenerative brake. This laboratory demonstrator validates the methodology developed to enrich the learning algorithm with physics.

The project demonstrated the feasibility of predicting noise and particulate emissions from friction braking systems, provided that data sensitive to the physical phenomena predominant in these emissions (mainly tribological and thermomechanical) are included. This innovative result opens the way to optimizing the use of friction braking systems, with the possibility of modulating the combination with regenerative braking on electric vehicles.
However, one constraint is the need to go through a learning phase on highly instrumented tests, which is costly and time-consuming. An improvement would be to be able to link global measurements available on industrial test benches (speed, force, etc.) with the data required for prediction. This can be achieved by learning methods explicitly integrating the physical phenomena linking the global parameters to those identified as key in the PI-CUBE project. Initial developments in the PI-CUBE project have proved promising. They show the need to include more in-depth physical modeling to describe the braking system (dynamic and thermomechanical response) and tribological mechanisms. It is also essential to combine these models with AI-based methods. This point is crucial, but also complex, as it tackles friction mechanisms that combine mechanical and physico-chemical phenomena on several scales, which are impossible to model explicitly to date. One prospect would be to add a mesoscopic dimension to the macroscopic description used in the PI-CUBE project, and link this modeling to the mesoscopic scale of material flows, with learning of the activation data for these flows. The need for experimental characterization would be greatly reduced to the data required to identify the learning algorithm associated with the physical modeling. A Franco-German ANR project (SLOPE) has been submitted with this objective.

Articles in peer-reviewed journals :
VV. LAI, I. PASZKIEWICZ, JF. BRUNEL, P. DUFRENOY (2022) Appearance of squeaks related to bearing surface tracking on a pin-on-disc system Mechanical Systems and Signal Processing 165, n°108364 (IF 6.82) DOI:10.1016/j.ymssp.2021.108364
C. GEIER, S. HAMDI, T. CHANCELIER, P. DUFRENOY, N. HOFFMANN, M. STENDER (2023) Machine learning-based state maps for complex dynamical systems: applications to friction-excited brake system vibrations Nonlinear Dynamics 111(24) 22137-22151 (IF 5.74) DOI: 10.1007/s11071-023-08739-6
3 submissions under review in MSSP, WEAR and LUBRICANTS

Conferences :
ENOC 2022 (10th European Nonlinear Dynamics Confeence), A machine learning perspective on frictional contacts and self-excited vibrations Maël Thévenot, Charlotte Geier, Jean-François Brunel, Merten Stender, Philippe Dufrénoy and Norbert Hoffmann
GAMM 2022, Data-driven stability maps for friction-induced vibrations Charlotte Geier, Merten Stender, Saïd Hamdi, Norbert Hoffmann, Thierry Chancelier
ASIABRAKE 2022 Identification of key braking emission parameters through model-experiment dialogue Philippe Dufrénoy, Yannick Desplanques, Jean-François Brunel, Mael Thevenot
CFM 2022 (Congrès Français de Mécanique) Tribology-thermomechanics relationship using model-experiment dialogues Philippe Dufrénoy (guest speaker)
EUROBRAKE 2023 Tribological analysis and link with particle emission from a pin-on-disc during braking. Maël Thévenot, Jean-François Brunel, Yannick Desplanques, Philippe Dufrénoy
JIFT 2023 Tribological analyses and link with noise and particle emissions of a pin-on-disc system during braking. Maël Thévenot, Jean-François Brunel, Yannick Desplanques, Philippe Dufrénoy, Merten Stender, Norbert Hoffmann
EUROBRAKE 2024 Analysis of multimodal experimental data for physical understanding of emission mechanisms during braking events. Maël Thévenot, Quentin Caradec, Jean-François Brunel, Nathanael Winter, Merten Stender, Norbert Hoffmann, Yannick Desplanques, Philippe Dufrénoy
LEEDS-LYON 2024 Multimodal experimental data analysis for the physical understanding of emission mechanisms during braking events. Maël Thévenot, Quentin Caradec, Jean-François Brunel, Florent Brunel, Merten Stender, Norbert Hoffmann, Philippe Dufrénoy

This project aims at developing a new braking control system for electric vehicles using artificial intelligence (AI) for a significant reduction of (i) brake particle emissions and of (ii) noise emissions. Both emissions are major sources of environmental pollution and health hazards in urban areas: Today there is consensus that today's vehicle friction brakes emit about the same amount of dust as present-day combustion engines. With electric drives, the friction brake (jointly with the tyres) will thus become the primary emitter of highly toxic dust from vehicles. A similar argument applies to noise: with most of the noise sources of combustion engines gone, the relative sliding of brake components will become even more audible and harmful than it is already today. Due to the complexity of the physics and chemistry involved in frictional braking, up to the present day, and despite decade long intense efforts, it has not been possible to generate satisfactory models, or a 'virtual twin', starting from fundamental theory and basic principles. Recent disruptive progress through AI methods suggests that a combination of physics and data-processing based methods is the way to move forward to applicable modelling, simulation and control for a greener and healthier mobility. With the paradigm change towards electric vehicles ahead, there now is the pressing need for obtaining a control of the future friction brakes with respect to cutting down emissions of particles and noise by intelligent use of braking. Having two means of braking (electric and mechanic) opens up a way of optimizing their combination. Different AI methods will be integrated: to enrich the physical models with a virtual sensor, to develop emission descriptors and to continuously improve the control strategy by learning with use. Models for predicting the load variables expected during the next braking event will determine the risk of a sharp increase in emissions in order to avoid critical operating points by using data-based control (learning by reinforcement) of electrical and mechanical brakes. The operation of the system will be demonstrated on a laboratory prototype. Hamburg University of Technology has a long-standing expertise on data science and computer modelling of noise emissions from friction brakes and is one of the leading teams in the field in Germany and world-wide. University of Lille, which has a long experience of friction interfaces and wear problems, is the leading European institution working on the tribology of friction brakes. Volkswagen AG and AUDI AG are Germany’s leading car manufacturers and allow access to real-world commercial brake testing and field data. Hitachi Automotive Systems is a world-leading friction brake manufacturer. All partners in total provide the complete knowledge, capability and excellence to develop this new AI based braking control strategy that will help industrial partners to develop new digital products, while research institutions will define future AI requirements for mechanical systems. In this way, the project represents the core of future interdisciplinary collaborations and technological developments with the aim of achieving environmentally friendly transport. The project is thus strengthening the European mobility industry, and at the same time pushing AI research for integration into mechanical systems.