Search Results

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.

Abstract Lane changes are stressful maneuvers for drivers, particularly during high-speed traffic flows. However, modeling driver’s lane-changing decision and implementation process is challenging due to the complexity and uncertainty of driving behaviors. To address this issue, this article presents a personalized Lane-Changing Model (LCM) for Advanced Driver Assistance System (ADAS) based on deep learning method. The LCM contains three major computational components. Firstly, with abundant inputs of Root Residual Network (Root-ResNet), LCM is able to exploit more local information from the front view video data. Secondly, the LCM has an ability of learning the global spatial-temporal information via Temporal Modeling Blocks (TMBs). Finally, a two-layer Long Short-Term Memory (LSTM) network is used to learn video contextual features combined with lane boundary based distance features in lane change events.

Abstract Obstacle avoidance is an important function in self-driving vehicle control. When the vehicle move from any arbitrary start positions to any target positions in environment, a proper path must avoid both static obstacles and moving obstacles of arbitrary shape. There are many possible scenarios, manually tackling all possible cases will likely yield a too simplistic policy. In this paper reinforcement learning is applied to the problem to form effective strategies. There are two major challenges that make self-driving vehicle different from other robotic tasks. Firstly, in order to control the vehicle precisely, the action space must be continuous which can’t be dealt with by traditional Q-learning. Secondly, self-driving vehicle must satisfy various constraints including vehicle dynamics constraints and traffic rules constraints. Three contributions are made in this paper.

Abstract This work introduces the concept of a dual-layer specification structure for standards that separate interoperability functions, such as backward compatibility, localization, and deployment, from those essential to reliability, security, and functionality. The latter group of features, which constitute the actual standard, make up the baseline layer for instructions, while all the elements required for interoperability are specified in a second layer, known as a Protocols, Operations, Usage, and Formats (POUF) document. We applied this technique in the development of a standard for Uptane [1], a security framework for over-the-air (OTA) software updates used in many automobiles. This standard is a good candidate for a dual-layer specification because it requires communication between entities, but does not require a specific format for this communication.

Abstract Modern vehicles integrate Internet of Things (IoT) components to bring value-added services to both drivers and passengers. These components communicate with the external world through different types of interfaces including the on-board diagnostics (OBD-II) port, a mandatory interface in all vehicles in the United States and Europe. While this transformation has driven significant advancements in efficiency and safety, it has also opened a door to a wide variety of cyberattacks, as the architectures of vehicles were never designed with external connectivity in mind, and accordingly, security has never been pivotal in the design. As standardized, the OBD-II port allows not only direct access to the internal network of the vehicle but also installing software on the Electronic Control Units (ECUs).

Abstract In an automated machining process, monitoring the conditions of the tool is essential for deciding to replace or repair the tool without any manual intervention. Intelligent models built with sensor information and machine learning techniques are predicting the condition of the tool with good accuracy. In this study, statistical models are developed to identify the conditions of the abrasive grinding wheel using the Acoustic Emission (AE) signature acquired during the surface grinding operation. Abrasive grinding wheel conditions are identified using the abrasive wheel wear plot established by conducting experiments. The piezoelectric sensor is used to capture the AE from the grinding process, and statistical features of the abrasive wheel conditions are extracted in time and wavelet domains of the signature. Machine learning algorithms, namely, Classification and Regression Trees (CART) and Support Vector Classifiers (SVC), are used to build statistical models.

Abstract Future SAE Level 4 and Level 5 autonomous vehicles (AV) will require novel applications of localization, perception, control, and artificial intelligence technology in order to offer innovative and disruptive solutions to current mobility problems. This article concentrates on low-speed autonomous shuttles that are transitioning from being tested in limited traffic, dedicated routes to being deployed as SAE Level 4 automated driving vehicles in urban environments like college campuses and outdoor shopping centers within smart cities. The Ohio State University has designated a small segment in an underserved area of the campus as an initial AV pilot test route for the deployment of low-speed autonomous shuttles. This article presents initial results of ongoing work on developing solutions to the localization and perception challenges of this planned pilot deployment.

Abstract To evaluate that automated vehicle is as safe as a human driver, a following question is studied: how does an automated vehicle react under extreme conditions close to collision? In order to understand the collision avoidance capability of an automated vehicle, we should analyze not only such post-extreme condition behavior but also pre-extreme condition behavior. We present a theory to analyze the collision avoidance capability of automated driving technologies. We also formulate a collision avoidance equation on the theory. The equation has two types of solutions: response driving plans and preparation driving plans. The response driving plans are supported by response strategy on which the vehicle reacts after detection of a hazard and they are highly efficient in terms of travel time.

Abstract Management of expressway networks has been mainly focused on defect management without looking at the correlations with accidental risks. This causes unsustainability in expressway infrastructure maintenance since such defects may not be a contributing factor toward public safety. Thus it is necessary to incorporate accidental events for decision-making in infrastructure management. This study has developed a novel approach to machine learning (ML) that incorporates actual primary data from the last 10 years of single-vehicle accidents (SVA) by collisions with motorway facilities, or so-called single-vehicle collisions with fixed objects. The ML is firstly aimed at identifying the influential factors of SVA in relation to finding effective countermeasures for accidents by integrating the correlation analysis, multiple regression analysis, and ML techniques. The study reveals that wet pavement conditions have a significant effect on SVA.

In this research, the interior noise of a passenger car was measured, and the sound quality metrics including sound pressure level, loudness, sharpness, and roughness were calculated. An artificial neural network was designed to successfully apply on automotive interior noise as well as numerous different fields of technology which aim to overcome difficulties of experimentations and save cost, time and workforce. Sound pressure level, loudness, sharpness, and roughness were estimated by using the artificial neural network designed by using the experiment values. The predicted values and experiment results are compared. The comparison results show that the realized artificial intelligence model is an appropriate model to estimate the sound quality of the automotive interior noise. The reliability value is calculated as 0.9995 by using statistical analysis.

Strict regulations on air pollution motivates clean combustion research for fossil fuels. To numerically mimic real gasoline fuel reactivity, surrogates are proposed to facilitate advanced engine design and predict emissions by chemical kinetic modelling. However, chemical kinetic models could not accurately predict non-regular emissions, e.g. aldehydes, ketones and unsaturated hydrocarbons, which are important air pollutants. In this work, we propose to use machine-learning algorithms to achieve better predictions. Combustion chemistry of fuels constituting of 10 neat fuels, 6 primary reference fuels (PRF) and 6 FGX surrogates were tested in a jet stirred reactor. Experimental data were collected in the same setup to maintain data uniformity and consistency under following conditions: residence time at 1.0 second, fuel concentration at 0.25%, equivalence ratio at 1.0, and temperature range from 750 to 1100K.

Particulate emissions from gasoline direct injection (GDI) engines continue to be a topic of substantial research interest. Forthcoming regulation both in the USA and the EU will further reduce their emission and drive innovation. Substantial research effort is spent undertaking experiments to understand, characterize, and research particle number (PN) emissions from engines and vehicles. Recent advances in computing power, data storage, and understanding of artificial intelligence algorithms now mean that these are becoming an important tool in engine research. In this work a random forest (RF) algorithm is used for the prediction of PN emissions from a highly boosted (up to 32 bar BMEP) GDI engine. Particle size, concentration, and the accumulation mode geometric standard deviation (GSD) are all predicted by the model. The results are analysed and an in depth study on parameter importance is carried out.

On-board diagnostics (OBD) data contain valuable information including real-world measurements of vehicle powertrain parameters. These data can be used to gain a richer data-driven understanding of complex physical phenomena like emissions formation during combustion. In this study, we develop a physics-based machine learning framework to predict and analyze trends in engine-out NOx emissions from diesel and diesel-hybrid heavy-duty vehicles. This model differs from black-box machine learning models presented in previous literature because it incorporates engine combustion parameters that allow physical interpretation of the results. Based on chemical kinetics and the characteristics of diffusive combustion, NOx emissions from compression ignition engines primarily depend non-linearly on three parameters: adiabatic flame temperature, the oxygen concentration in the cylinder when the intake valves are closed, and combustion time duration.

Even under consideration of the increasing dynamics of measures to reduce CO2 in the transport sector and the resulting, now visible changes in the development and registration of new passenger cars (electrification), it is anticipated that vehicle drives containing an internal combustion engine will continue to have significant market shares in the medium to long term. It is assumed that a significant proportion of these cars will be hybrid vehicles in the future. As a result, in order to implement future requirements for improved air pollution control (Post EU6/Zero Impact) and CO2 reduction, consideration of these aspects must be an integral part of the application and of development activities in general. At TU Darmstadt, a consistent method for the development of powertrains with regard to their relevant real-world driving emissions - the Most Relevant Testing Procedure (MRTP), was established.

Performance predictions of advanced turbocharged engines are becoming difficult because conventional engine models are built using performance map data of turbochargers with a proportional integral derivative (PID) controller. Improving prediction capabilities under transient test cycles or real driving conditions is a challenging task. This study applies a machine learning technique to predict turbocharger performances with high accuracy under steady-state and transient conditions. The manipulated signals of engine speed and torque created based on Compressed High-Intensity Radiated Pulse (Chirp signal) and Amplitude-modulated Pseudo-Random Binary Signal (APRBS) are used as inputs to the engine testbed. Data from the engine experiments are used as training data for the AI-based turbocharger model. High prediction accuracy of the AI turbocharger model is achieved with the co-efficient of determination in the model, and cross-validation results are higher than 0.8.

Due to constant evolution in both noise regulations and noise comfort standards, noise reduction inside the vehicle remains one of the main issues faced today by the automotive industry. One of the most efficient methods for noise reduction is the introduction of acoustic treatments, made of multilayered trimmed panels. Constraints on these components, such as weight, packaging space and overall sound quality as well as the amount of possible material and geometrical combinations, have led automotive OEMs to use innovative methods, such as numerical acoustic simulation, so as to evaluate noise transmission in a fast and cost-effective way. While the computational cost for performing such analyses is insignificant for a limited number of configurations, the evaluation of multiple design parameter combinations early in the design stage can lead to non-viable computation times in an industrial context.

As climate change looms large, the automotive industry gears up for an Electric Vehicle (EV) transition to pull down our net global greenhouse emissions to zero together with the clean energy transition. It becomes the need of the hour to optimize the use of our resources and meet the requirements of time, effort, cost, accuracy and transient performance brought in by the stringent emission norms and the Real Driving Emissions (RDE) test. The authors present a Reinforcement learning technique to address the real-world challenges for accelerated product development. Reinforcement Learning was used to parameterize a time varying electromechanical system and proved effective in modelling the stochastic nature of processes in powertrain development.

As per the 2018 MoRTH accident report, there were 467,044 accidents, out of which 137,726 were fatal which resulted in 151,417 fatalities. In order to get an idea of the reasons for injuries and estimate the benefits of any intervention, a mathematical model should go a long way. This study is aimed at the development of such a model to predict the injuries sustained by the occupants of an M1 vehicle. We used a detailed accident database of 'Road Accident Sampling System India' (RASSI). RASSI, since 2011, has been collecting traffic accident data scientific across various locations in India. In the data, the occupant injuries are classified as No injury, Minor, Serious and Fatal We used the data of about 4700+ M1 occupants for the study & used almost 40 input parameters to determine the outcome. Based on the data, an algorithm was developed with an overall accuracy of about 67%. The parameters represented human, infrastructure, and environment.

In an automotive braking system, Vacuum pump is used to generate vacuum in the vacuum servo or brake booster in order to enhance the safety and comfort to the driver. The vacuum pump operation in the braking system varies from conventional to electric vehicles. The vacuum pump is connected to the alternator shaft or CAM shaft in a conventional vehicle, operates continuously at engine speed and supplies continuous vacuum to the brake servo irrespective of vacuum requirement. To sustain continuous operation, these vacuum pumps are generally oil cooled. Whereas in electric vehicles, the use of a motor-driven vacuum pump is very much needed for vacuum generation as there is no engine present. Thus, with the assistance of an electronic control unit (ECU), the vacuum pump can be operated only when needed saving a significant amount of energy contributing to fuel economy and range improvement and emission reduction.

The recent development pace of the automotive technology is so rapid worldwide. Especially in a green car, hybrid electric vehicles (HEVs) have been studied a lot due to their significant effects on the urban driving. In the vehicle energy management strategy study, the driving speed is assumed to be known in advance, however the speed is not given in a real world. Accordingly, the prediction of vehicle speed is very important. In this study, we study the prediction methodology for the speed prediction using deep learning. Based on the vehicle driving speed data, the supervised deep learning has been used and the speed prediction accuracy using deep learning shows accurate results comparing to the actual speed. The supervised deep learning is used which is suitable for driving cycle database. As a result, the speed prediction after few seconds is feasible.