Search Results

Convolutional neural networks are the de facto method of processing camera, radar, and lidar data for use in perception in ADAS and L4 vehicles, yet their operation is a black box to many engineers. Unlike traditional rules-based approaches to coding intelligent systems, networks are trained and the internal structure created during the training process is too complex to be understood by humans, yet in operation networks are able to classify objects of interest at error rates better than rates achieved by humans viewing the same input data.

Driving simulators allow the testing of driving functions, vehicle models and acceptance assessment at an early stage. For a real driving experience, it's necessary that all immersions are depicted as realistically as possible. When driving manually, the perceived haptic steering wheel torque plays a key role in conveying a realistic steering feel. To ensure this, complex multi-body systems are used with numerous of parameters that are difficult to identify. Therefore, this study shows a method how to generate a realistic steering feel with a nonlinear open-loop model which only contains significant parameters, particularly the friction of the steering gear. This is suitable for the steering feel in the most driving on-center area. Measurements from test benches and real test drives with an Electric Power Steering (EPS) were used for the Identification and Validation of the model.

Modeling thermal systems in Battery Electric Vehicles (BEVs) is crucial for enhancing energy efficiency through predictive control strategies, thereby extending vehicle range. A major obstacle in this modeling is the often limited availability of detailed system information. This research introduces a methodology using neural networks for system identification, a powerful technique capable of approximating the physical behavior of thermal systems with minimal data requirements. By employing black-box models, this approach supports the creation of optimization-based operational strategies, such as Model Predictive Control (MPC) and Reinforcement Learning-based Control (RL). The system identification process is executed using MATLAB Simulink, with virtual training data produced by validated Simulink models to establish the method's feasibility. The neural networks utilized for system identification are implemented in MATLAB code.

The calibration of Engine Control Units (ECUs) for road vehicles is challenged by stringent legal and environmental regulations, coupled with short development cycles. The growing number of vehicle variants, although sharing similar engines and control algorithms, requires different calibrations. Additionally, modern engines feature increasingly number of adjustment variables, along with complex parallel and nested conditions within the software, demanding a significant amount of measurement data during development. The current state-of-the-art (White Box) model-based ECU calibration proves effective but involves considerable effort for model construction and validation. This is often hindered by limited function documentation, available measurements, and hardware representation capabilities. This article introduces a model-based calibration approach using Neural Networks (Black Box) for two distinct ECU functional structures with minimal software documentation.

Autonomous Emergency Braking (AEB) systems are critical in preventing collisions, yet their effectiveness hinges on accurately estimating the distance between the vehicle and other road users, as well as understanding road conditions. Errors in distance estimation can result in premature or delayed braking and varying road conditions alter road-tire friction coefficients, affecting braking distances. Advancements in sensor technology and deep learning have improved vehicle perception and real-world understanding. The integration of advanced sensors like LiDARs has significantly enhanced distance estimation. Cameras and deep neural networks are also employed to estimate the road conditions. However, AEB systems face notable challenges in urban environments, influenced by complex scenarios and adverse weather conditions such as rain and fog. Therefore, investigating the error tolerance of these estimations is essential for the performance of AEB systems.

Investigating human driver behavior enhances the acceptance of the autonomous driving and increases road safety in heterogeneous environments with human-operated and autonomous vehicles. The previously established driver fingerprint model, focuses on the classification of driving style based on CAN bus signals. However, driving styles are inherently complex and influenced by multiple factors, including changing driving environments and driver states. To comprehensively create a driver profile, an in-car measurement system based on the Driver-Driven vehicle-Driving environment (3D) framework is developed. The measurement system records emotional and physiological signals from the driver, including ECG signal and heart rate. A Raspberry Pi camera is utilized on the dashboard to capture the driver's facial expressions and a trained convolutional neural network (CNN) recognizes emotion. To conduct unobtrusive ECG measurements, an ECG sensor is integrated into the steering wheel.

Autonomous driving is a hot topic in the automotive domain, and there is an increasing need to prove its reliability. They use machine learning techniques, which are themselves stochastic techniques based on some kind of statistical inference. The occurrence of incorrect decisions is part of this approach and often not directly related to correctable errors. The quality of the systems is indicated by statistical key figures such as accuracy and precision. Numerous driving tests and simulations in simulators are extensively used to provide evidence. However, the basis of all descriptive statistics is a random selection from a probability space. The difficulty in testing or constructing the training and test data set is that this probability space is usually not well defined. To systematically address this shortcoming, ontologies have been and are being developed to capture the various concepts and properties of the operational design domain.

In recent years, the urgent need to fully exploit the fuel economy potential of the Electrified Vehicles (xEVs) through the optimal design of their Energy Management System (EMS) have led to an increasing interest in Machine Learning (ML) techniques. Among them, Reinforcement Learning (RL) seems to be one of the most promising approaches thanks to its peculiar structure, in which an agent is able to learn the optimal control strategy through the feedback received by a direct interaction with the environment. Therefore, in this study, a new Soft Actor-Critic agent (SAC), which exploits a stochastic policy, was implemented on a digital twin of a state-of-the-art diesel Plug-in Hybrid Electric Vehicle (PHEV) available on the European market. The SAC agent was trained to enhance the fuel economy of the PHEV while guaranteeing its battery charge sustainability.

Traditional CACC systems utilize inter-vehicle wireless communication to maintain minimal yet safe inter-vehicle distances, thereby improving traffic efficiency. However, introducing communication delays generates system uncertainties that jeopardize string stability, a crucial requirement for robust CACC performance. To address these issues, we introduce a decentralized Model Predictive Control (MPC) approach that incorporates Kalman Filters and state predictors to counteract the uncertainties posed by noise and communication delays. We validate our approach through MATLAB Simulink simulations, using stochastic and mathematical models to capture vehicular dynamics, Wi-Fi communication errors, and sensor noises. In addition, we explore the application of a Reinforcement Learning (RL)-based algorithm to compare its merits and limitations against our decentralized MPC controller, considering factors like feasibility and reliability.

In vehicle NVH development, vibroacoustic simulations with Finite Element (FE) models are a common technique. The computational costs for these calculations are steadily rising due to more detailed modelling and higher frequency ranges. At the same time, the need for multiple evaluations of the same model with different input parameters, e.g., for uncertainty quantification, optimization, or robustness investigations, is also increasing. Therefore, it is crucial to reduce the computational costs in these cases. A common technique is to use surrogate models that replace the computationally intensive FE model to perform repeated evaluations. Several different methods in this area are well established, but with the continuous advancements in the field of machine learning, interesting new methods like the Gaussian Process (GP) regression arises as a promising approach.

This research aims presents the method classifying the noise source and evaluating the sound quality of the noise caused by operating of electric power steering wheel in an electric vehicle. The steering wheel has been operated by the motor drive by electric power and it called motor-driven electric power (MDPS) system. If the motor is attached to the steering column of the steering device, it is called C-MDPS system. The steering device of the C-MDPS system comprises of motor, bearings, steering column, steering wheel and worm shaft. Among these components the motor and bearings are main noise sources of C-MDPS system. When the steering wheel is operated in an electric vehicle, the operating noise of the steering device inside the vehicle is more annoying than that in a gasoline engine vehicle since the operating noise is not masked by engine noise. Defects in the C-MDPS system worsen the operating noise of the steering system.

Though modal analysis is a common tool to evaluate the dynamic properties of a structure, there are still many individual decisions to be made during the process which are often based on experience and make it difficult for occasional users to gain reliable and correct results. One of those experience-based choices is the correct number and placement of reference points. This decision is especially important, because it must be made right in the beginning of the process and a wrong choice is only noticeable in the very end of the process. Picking the wrong reference points could result in incomplete modal analysis outcomes, as it might make certain modes undetectable, compounded by the user's lack of awareness about these missing modes. In the paper an innovative approach will be presented to choose the minimal number of mandatory reference points and their placement.

The high-frequency whining noise produced by motors in modern electric vehicles causes a significant issue, leading to annoyance among passengers. This noise becomes even more noticeable due to the quiet nature of electric vehicles, which lack other noises to mask the high-frequency whining noise. To improve the noise caused by motors, it is essential to optimize various motor design parameters. However, this task requires expert knowledge and a considerable time investment. In this study, we explored the application of artificial intelligence to optimize the NVH performance of motors during the design phase. Firstly, we selected and modeled three benchmark motor types using Motor-CAD. Machine learning models were trained using Design of Experiment methods to simulate batch runs of Motor-CAD inputs and outputs.

Liebherr Machines Bulle SA designs and produces High-quality diesel engines, injection systems as well as hydraulic components. Liebherr has an Acoustic End of Line (A-EoL) system on serial test benches. All engines are measured, and noises are evaluated by operators. This subjective evaluation leads to dispersion on the evaluations, particularly for whining noise. To achieve Swiss quality requirements and ensure customer satisfaction, Liebherr wishes to define a new methodology to find a quantitative and objective criterion to set a robust engine noise compliance standard. This new methodology is based on near field microphone measurement of an engine run-down. First, whining noise signatures are extracted from the raw signal. Secondly, psychoacoustic indicators are calculated on the isolated signatures. Thresholds are then established to validate engine deliveries.

The engine acoustic character has always represented the product DNA, owing to its strong correlation with in-cylinder pressure gradient, components design and perceived quality. Best practice for engine acoustic characterization requires the employment of a hemi-anechoic chamber, a significant number of sensors and special acoustic insulation for engine ancillaries and transmission. This process is highly demanding in terms of cost and time due to multiple engine working points to be tested and consequent data post-processing. Since Neural Networks potentially predicting capabilities are apparently un-exploited in this research field, the following paper provides a tool able to acoustically estimate engine performance, processing system inputs (e.g. Injected Fuel, Rail Pressure) thanks to the employment of Multi Layer Perceptron (MLP, a feed forward Network working in stationary points).

Electrified vehicles are particularly quiet, especially at low speeds due to the absence of combustion noises. This is why there are laws worldwide for artificial driving sounds to warn pedestrians. These sounds are generated using a so-called Acoustic Vehicle Alerting System (AVAS) which must maintain certain minimum sound pressure levels in specific frequency ranges at low speeds. The creation of the sound currently involves an iterative and sometimes time-consuming process that combines composing the sound on a computer with measuring the levels with a car on an outside noise test track. This continues until both the legal requirements and the subjective demands of vehicle manufacturers are met. To optimize this process and reduce the measurement effort on the outside noise test track, the goal is to replace the measurement with a simulation for a significant portion of the development.

Acoustic palliatives used in the automotive industry have evolved from simple felt and heavy layer combinations into highly complex formulations and combinations to account for higher performance targets, lower weight and inevitably cost constraints. Achieving Customer performance compliance usually involves a time-consuming exercise of material characterisation and measurement. Ideally this should be carried out via simulation, but as material mixtures and compositions become more complex, the ability to accurately simulate their acoustic performance is becoming increasingly difficult. Historically, Biot parameters and their associated TMM models have been used to simulate the acoustic performance of multi-layer material compositions. However, these simulations are not able to account for real-world complexities such as manufacturing imperfections or inter-layer gluing effects.

Trim materials are often used for vibroacoustic energy absorption purposes within vehicles. To estimate the sound impact at a driver’s ear, the substructuring approach can be applied. Thus, transfer functions are calculated starting from the acoustic source to the car body, from the car body to the trim and, finally, from the trim to the inner cavity where the driver is located. One of the most challenging parts is the calculation of the transfer functions from the car body inner surface to the bottom trim surface. Commonly, freely laying mass-spring systems (trims) are simulated with a fixed boundary and interface phenomena such as friction, stick-slip or discontinuities are not taken into consideration. Such an approach allows for faster simulations but results in simulations strongly overestimating the energy transfer, particularly in the frequency range where the mass-spring system’s resonances take place.

The course will enable the learner to apply the fundamental principles behind safety of machine learning to a wide range of applications. The course guides learners through an appropriate selection of methods and tools tailored to the learner’s specific projects. With the acquired knowledge the learner will be able to shape the development and assessment of ML-based safety-related functions enabling their teams to leverage the power of advanced ML techniques without undermining safety.

Landing of spacecraft on Lunar or Martian surfaces is the last and critical step in inter planetary space missions. The atmosphere on earth is thick enough to slow down the craft but Moon or Mars does not provide a similar atmosphere. Moreover, other factors such as lunar dust, availability of precise onboard navigational aids etc would impact decision making. Soft landing meaning controlling the velocity of the craft from over 6000km/h to zero. If the craft’s velocity is not controlled, it might crash. Various onboard sensors and onboard computing power play a critical role in estimating and hence controlling the velocity, in the absence of GPS-like navigational aids. In this paper, an attempt is made using visual onboard sensor to estimate the velocity of the object. The precise estimation of an object's velocity is a vital component in the trajectory planning of space vehicles, particularly those designed for descent onto lunar or Martian terrains, such as orbiters or landers.