Riassunto analitico
The friction coefficient is of high importance to the automotive sector. Vital components such as brakes, clutches and tires depend on the friction coefficient for their principal functionality. A more accurate description of this coefficient will ultimately lead to an increase in vehicle performance and a more efficient way of managing the car concept phase. Mapping the pressure-torque relationship, in a dynamic model, would substantially reduce the development costs of advanced vehicle systems by increasing the fidelity of inexpensive software, hardware, or model in the loop systems. In addition, the introduction of a reliable friction coefficient in thermal models allows simple quantification of thermal energy, based of pressure, velocity, and temperature telemetry, thereby decreasing testing and development cost, and increasing the reliability and potency of the model. The study targets to obtain a reliable description of the friction coefficient for a wide range of operating conditions. The model is validated against real world data as well as laboratory-controlled brake dynamometer data, using specifically designed tests which ensure adequate quality and quantity of data. From literature it follows that friction is predominantly influenced by the applied normal force (or pressure), relative velocity and surface temperature, however, it is widely accepted that not all influencing factors have been identified, which consequently impedes a correct detailed analytic expression of the coefficient. Moreover, some of influencing factors are difficult to determine without the usage of time and computationally intensive Finite Element Modelling. The approach of this study aims to circumvent these challenges by applying machine learning, specifically a Random Forest regression model generated using MATLAB. Effectively the model does not need a qualitative description of the behaviour of friction, thereby bypassing the analytical indeterminism. By using an ensemble of simple models, as per the Random Forest framework, called Bootstrap Aggregation, the computational simplicity and the prevention of overfitting is ensured. The intended use of the model is to provide valuable information regarding the friction coefficient for the determination of the geometry of the braking system, performance, and cooling targets throughout the concept and development phase of a brake system of an automotive vehicle. Bench tests are conducted on the brake system as soon as the materials and the rough geometry of the friction couple are determined. The test results are used to train the model, hereby characterizing the friction behaviour for a wide range of operating conditions and allowing the individual contribution of each input parameter to be quantified. During the concept phase, the model acts as an analysis tool, whereby individual inputs are related to the friction behaviour and can be compared between friction couples. After the concept phase, the model can be implemented as predictive instrument and its accuracy can be increased with any additional tests. The results demonstrate that the complex behaviour of a highly variable friction coefficient can be accurately predicted for a wide range of real-world data using a machine learning model trained with brake dynamometer data. Although the model has shown to have an adequate accuracy when using only pressure, velocity and temperature as inputs, the model is also compatible with more inputs, e.g. acceleration, hereby increasing its potency as an analysis tool and increasing its accuracy as predictive tool. The friction behaviour can be analysed for any response independently of the manoeuvre, notably predicting the deviant first brake event with a high accuracy and is not limited to one vehicle.
|
Abstract
The friction coefficient is of high importance to the automotive sector. Vital components such as brakes, clutches and tires depend on the friction coefficient for their principal functionality. A more accurate description of this coefficient will ultimately lead to an increase in vehicle performance and a more efficient way of managing the car concept phase. Mapping the pressure-torque relationship, in a dynamic model, would substantially reduce the development costs of advanced vehicle systems by increasing the fidelity of inexpensive software, hardware, or model in the loop systems. In addition, the introduction of a reliable friction coefficient in thermal models allows simple quantification of thermal energy, based of pressure, velocity, and temperature telemetry, thereby decreasing testing and development cost, and increasing the reliability and potency of the model.
The study targets to obtain a reliable description of the friction coefficient for a wide range of operating conditions. The model is validated against real world data as well as laboratory-controlled brake dynamometer data, using specifically designed tests which ensure adequate quality and quantity of data.
From literature it follows that friction is predominantly influenced by the applied normal force (or pressure), relative velocity and surface temperature, however, it is widely accepted that not all influencing factors have been identified, which consequently impedes a correct detailed analytic expression of the coefficient. Moreover, some of influencing factors are difficult to determine without the usage of time and computationally intensive Finite Element Modelling.
The approach of this study aims to circumvent these challenges by applying machine learning, specifically a Random Forest regression model generated using MATLAB. Effectively the model does not need a qualitative description of the behaviour of friction, thereby bypassing the analytical indeterminism. By using an ensemble of simple models, as per the Random Forest framework, called Bootstrap Aggregation, the computational simplicity and the prevention of overfitting is ensured.
The intended use of the model is to provide valuable information regarding the friction coefficient for the determination of the geometry of the braking system, performance, and cooling targets throughout the concept and development phase of a brake system of an automotive vehicle. Bench tests are conducted on the brake system as soon as the materials and the rough geometry of the friction couple are determined. The test results are used to train the model, hereby characterizing the friction behaviour for a wide range of operating conditions and allowing the individual contribution of each input parameter to be quantified. During the concept phase, the model acts as an analysis tool, whereby individual inputs are related to the friction behaviour and can be compared between friction couples. After the concept phase, the model can be implemented as predictive instrument and its accuracy can be increased with any additional tests.
The results demonstrate that the complex behaviour of a highly variable friction coefficient can be accurately predicted for a wide range of real-world data using a machine learning model trained with brake dynamometer data. Although the model has shown to have an adequate accuracy when using only pressure, velocity and temperature as inputs, the model is also compatible with more inputs, e.g. acceleration, hereby increasing its potency as an analysis tool and increasing its accuracy as predictive tool. The friction behaviour can be analysed for any response independently of the manoeuvre, notably predicting the deviant first brake event with a high accuracy and is not limited to one vehicle.
|