Autonomous driving and other megatrends bring disruptive change to the automotive industry. In order to cope with this change in the field of autonomous driving, BMW Group and its partners jointly developed a nonproprietary platform. This talk gives you insight into the main challenges of autonomous driving and points out how data-driven development and artificial intelligence are the keys to success.
Forty years ago, there were nearly no semiconductors in cars. Besides a generator that supplied power for lighting, ignition, and simple relays and switches, the car was a pure mechanical system. Today, the car is dominated by electronics and 80% of the innovation happening is based on semiconductor components.
We stand on the border of e-mobility, advanced driver assistance systems, and connected cars. We have the lowest level of fatalities on our roads, despite producing 80 million cars every year. On the other hand, the number of cars, specifically in fast-growing megacities, will require more modern mobility concepts, which will also have a strong impact on the car industry. The flexible and integrated mobility concept will promote car sharing, autonomous driving, connected car technologies, and new types or setups of cars from three-wheelers to small cars driving in convoy.
The semiconductor industry will be challenged to support new and demanding requirements from the automotive industry to enable self-driving. They’ll be asked to provide semiconductors with high computing power, high-speed networks, and redundant features to provide functional and safe operational systems. This will take several steps of learning, system optimization, and new semiconductor technology that is not used today in the automotive industry. In consequence, this dramatically increased complexity has to be managed along the automotive supply chain with new and more integrated development approaches, modern design and simulation tools, and the necessary verification processes to achieve the safety and security levels expected of the safe and efficient mobility systems of tomorrow.
The automobile went through its digital transformation with the addition of electronic controls in virtually every system. With automated driving and predictive maintenance, the automobile is experiencing another digital transformation in which data-driven algorithms for implementing artificial intelligence are playing a key role. In this presentation, Roy Lurie, who leads the development of MATLAB®, shares advances in MATLAB for handling big engineering data, making analytics and deep learning easy and accessible.
Autonomous driving, electrification, car-to-car and end-to-end communications as well as smart assistant systems are only some of the focus topics of new and current products in the automotive industry. All these products show a clear trend: They are no longer systems of one single domain but become more and more "systems of systems". In order to cope with the increasing complexity, different methods and technologies for model-based system engineering are available on the market. Furthermore, these new products include a steadily growing part of software components as their essential elements. These software components are integrated and implemented in a changing hardware architecture which needs high optimization of code and resource assignment. For this optimization, Model-Based Design, including automatic code generation, is already well established throughout the automotive industry in the areas of software design and software implementation. Nevertheless, topics like “time to market” and automation in general are still challenging existing processes and methods. That’s why, in the software industry, agile methods have gathered huge attention. Additionally, the increasing complexity and tighter integration of future product needs also require paying strong attention to safety and security aspects, and with this, existing and upcoming standards and regulations.
In this talk, you will hear about the results reached at Continental Automotive Division Powertrain throughout the last 15 years with their tool environment for Model-Based Design. In the second part of the talk, Marco Kunze will show how the company intends to close the gap between Model-Based system engineering and Model-Based Design using a seamless tool chain for Model-Based development. Providing full traceability and fulfilling safety standards are base requirements for the tool chain as well as support of agile development methods not only for software construction, but also at the early stage of requirement engineering and system architectural design.
Marco Kunze, Continental Automotive
At Bosch, an in-house developed program is used for engineering push belts—a critical part of the continuous variable transmission (CVT). The main function of the program is to estimate product lifetime by calculating effective material stresses and comparing these to available SN data. To provide worldwide access to this program, a web-based version was designed, which uses a SQL database to store all inputs and outputs, a web interface (Tomcat®), business logic written in Java®, and MATLAB® to program the engineering calculations. In order to allow many users to access the system concurrently, MATLAB Production Server™ is used. Requests for calculations from the Java part of the program are sent to the input queue of MATLAB Production Server. The duration of the calculations can be anything between one second and 24 hours, depending on the size and complexity of the calculation task. There are many benefits of this design: no installation of any software on users' computers, reduced costs for software, a clean separation between the components of the system, and short response times if bug fixes or improvements to the MATLAB code is required.
Dirk Twisk, Bosch Transmission Technology
This presentation first defines a software development approach based on four types of design models used at four different phases of the software development lifecycle. Then, a specific quality objective, named Model Quality Objectives (MQO), is proposed for each type of model. Each objective is defined as a set of quality characteristics with some measurable criteria named Model Quality Requirement (MQR). Some additional guidelines are provided to manage the planning and quality assessment activities related to MQO and MQR. This presentation concludes with some expected impact on the adoption of MQO by the automotive industry and the preliminary results on its deployment.
Florian Levy, Renault
Stéphane Louvet, Robert Bosch
Francois Guerin, MathWorks
In modern vehicles, more and more different types of sensors come into play. Aside from radar, ultrasonic, and camera, the most recent developments deal with distance measurements using LiDAR sensors, where the distance is calculated using a time of flight measurement of directed laser beams. This new sensor type requires a new set of algorithms to be implemented for the automotive industry.
Johannes Michael, Valeo
In this presentation, you’ll see a demonstration of the automatic classification of vehicle data, which results from different driving maneuvers and aims to predict the oversteering of a vehicle. A comparison between a classical implementation and the machine learning approach is established. To date, certain thresholds have been defined in the classical sense, which should signal an overshoot when this threshold is exceeded. In some cases, there are also speed-dependent sleepers determined by many years of experience. Thus, the question of how well and how quickly measurements can be characterized using machine learning methods also comes into question.
A model based on a classification algorithm in Statistics and Machine Learning Toolbox™ was trained with a focus on the concrete driving situation of the overdrive. For the data sets used for training, the prediction accuracy of the model is over 95%. In the next step, this model is then applied to new records. The first evaluations show promising results.
Tobias Freudling, BMW Group
The requirements in terms of harmonizing disparate engine characteristics are ever more demanding. Power, efficiency, dynamic response, and exhaust emission quality are just four factors that need to be coordinated. To meet those requirements, engine functions and the associated calibration of the engine control system are increasingly complex. This applies in particular to one fundamental component of engine control: the air charge model. Without the use of powerful application tools, its exact calibration is no longer feasible. Porsche Engineering has developed an alternative method employing a Model-Based approach.
One of the great challenges of the air charge model is that the characteristics and maps have to be calibrated very precisely, although their outputs do not correspond to any directly measurable physical and thermodynamic values. Due to the interplay between the calculated values of the engine control unit and the complexity of calculation models, it has now become impossible to parameterize the air charge model—in other words, directly adjust the maps—during ongoing operations on the engine test bench. The basic calibration must therefore be conducted using special tools that enable correct calibration of maps using measurement data. The significance and use of such tools has risen enormously and their development is a core competence in the field of engine calibration at Porsche Engineering.
By inverting the logical path of a complete function whose outcome corresponds to a measurable physical variable, it is possible to derive the precise value of the map output for each executed operating point. As soon as the map outputs for all operating points are known, numerical models are created that calculate the relationships between the input and output values of the maps. Using these models, the respective maps are then calibrated.
This alternative approach has proven successful and is used for the calibration of almost all ECU models.
Matteo Skull, Porsche Engineering Services
In this presentation, see an efficient method for real-time monitoring of a large number of process parameters and its application to an example of real-life production data.
In modern production facilities, huge amounts of relevant process data are recorded. In series production, identical production cycles are repeated, producing data packages with characteristic time series for each cycle and process parameter. In normal production mode, the time series of each cycle are highly similar to each other. Deviations indicate malfunction or abrasion.
Within this presentation, the time series of a large number of process parameters are analyzed in order to detect deviations from an ideal production mode. A challenge of this task is the large variety and complexity of the features in the time series. In the training phase, the presented method reduces complex time series to a small number of characteristic numbers and their typical variations. In the monitoring phase, the characteristic numbers of each cycle are compared to their ideal values determined during training.
The characteristic numbers allow you to distinguish between deviations in time and signal shape. Deviations in time may arise due to delays or process conditions. Deviations in shape, such as changes in the amplitude values, may occur due to adjustments of machine parameters, malfunctions, or wear conditions.
The key benefit of the presentation is a method that characterizes complex signals by a small amount of characteristic numbers, and thus, can easily identify statistically significant deviations.
Jessica Fisch, Daimler
Irina Ostapenko, Dr. Türck Ingenieurbüro
This presentation demonstrates how to implement engineering data analytics applications quickly and efficiently with MATLAB®. It features a real data set obtained from real cars.
- Handling big data efficiently with MATLAB: importing and preprocessing
- Developing analytics for vehicle test fleet data
- Automatically detecting events of interest and computing key metrics such as brake-specific fuel consumption
- Scaling your application to clusters and clouds
Christoph Stockhammer, MathWorks
In this presentation, you will learn about the latest Simulink® tools for full vehicle simulation. With a library of powertrain components and complete vehicle models for simulating fuel economy, performance, and driving maneuvers, these tools give you a powerful platform for full vehicle simulation. You will see a demonstration of how these tools can be used to analyze energy management, characterize ride and handling, develop powertrain and chassis controls, and much more.
Mike Sasena, MathWorks
Lars Krause, MathWorks
Learn how to use MATLAB® for designing, developing, and deploying computer vision and deep learning applications on NVIDIA® Tesla® GPUs or Tegra® system-on-chips, whether on your local machine, in a cluster, or on embedded systems, including NVIDIA Jetson™ TK1/TX1/TX2 and DRIVE™ PX platforms. The workflow starts with algorithm design in MATLAB. The deep learning network is defined in MATLAB and is trained using GPU and parallel computing support for MATLAB, either on the desktop computer, a local compute cluster, or in the cloud. Then, the trained network is augmented with traditional computer vision techniques and the application is verified in MATLAB. Finally, a compiler automatically generates portable and highly optimized CUDA® code from the MATLAB algorithm, which is then implemented on the Tegra platform using cross-compilation. The execution speed of the auto-generated CUDA code is ~2.5x faster than Apache MXNet™, ~5x faster than Facebook Caffe2, ~7x faster than Google™ TensorFlow™, and comparable to an optimized TensorRT™ implementation.
Alexander Schreiber, MathWorks
Advanced driver assistance systems (ADAS) and autonomous driving technologies are redefining the automotive industry, changing all aspects of transportation, from daily commutes to long-haul trucking. Engineers across the industry use Model-Based Design with MATLAB® and Simulink® to develop their automated driving systems. This presentation demonstrates how MATLAB and Simulink serve as an integrated development environment for the different domains required for automated driving, including perception, sensor fusion, and control design.
MathWorks engineers demonstrate new technologies to accelerate development of ADAS and automated driving applications with MATLAB and Simulink, including:
- Designing vision detection algorithms with deep learning
- Designing sensor fusion algorithms with recorded and live data
- Designing control algorithms with model predictive control
Gaurav Tomar, MathWorks
Marco Roggero, MathWorks