Powertrain Development with Artificial Intelligence: History, Work Processes, Concepts, Methods and Application Examples

Description

The variety of future powertrain concepts has drastically increased the development cost for automotive manufactures. Profitable investment requires a significantly leaner and efficient powertrain development process. Traditional methods of test and model based development need to be assisted by big data and data analytics. For this purpose, a valuable tool is available at the right time - artificial intelligence (AI). But what does AI really mean in a narrower sense? What concepts lie behind it? And how are the methods and algorithms transferable to powertrain applications? For the first time, this book aims to bridge the gap between automotive engineering and computer science, by illuminating the complexity of current AI concepts and demystifying it for powertrain applications. By elaborating on work processes, it shows how AI could be implemented and how completely novel methods can help us reshape the future of mobility. Contents List of Figures List of Tables 1 Introduction 2 The Combustion Engine at the Turn of Industrialization 2.1 The Prehistory 2.2 The Technological Revolution 3 Revolution through Simulation for the Development of Powertrains 3.1 Flow Calculation 3.1.1 3D-CFD 3.1.2 1D-CFD 3.1.3 0D-CFD 3.2 Elastohydrodynamics (EHD) and Tribology 3.3 Combustion Chamber (Combustion and Thermodynamics) 3.4 Material Strength and Structure 3.4.1 Theory of Elasticity 3.4.2 Alternative Methods 3.5 Acoustics 4 Big Data for Powertrains 4.1 The Future of Powertrain Concepts 4.2 The History of Simulation 4.3 Development Processes and Scenarios of Simulation with Big Data 5 Powertrain Development with Artificial Intelligence [DELETE] 5.1 Models of Artificial Intelligence 5.2 Overview: Levels of AI 5.3 Search Tree 5.4 Machine Learning 5.4.1 Supervised Learning 5.4.2 Unsupervised Learning 5.4.3 Reinforcement Learning 5.5 Artificial Neural Networks (ANN) 5.5.1 Activation Function 5.5.2 Feedforward Networks (ANN) and Recurrent Networks (RNN) 5.5.3 Training Procedure 5.5.4 Network Architecture and Performance 5.5.5 Fuzzy Logik 5.6 Deep Learning 5.6.1 Convolutional Neural Network (CNN) 5.6.2 Generative Adversarial Network (GAN) 5.7 Software Bibliography 1 Introduction 2 The Combustion Engine at the Turn of Industrialization 2.1 The Prehistory 2.2 The Technological Revolution 3 Revolution through Simulation for the Development of Powertrains 3.1 Flow Calculation 3.1.1 3D-CFD 3.1.2 1D-CFD 3.1.3 0D-CFD 3.2 Elastohydrodynamics (EHD) and Tribology 3.3 Combustion Chamber (Combustion and Thermodynamics) 3.4 Material Strength and Structure 3.4.1 Theory of Elasticity 3.4.2 Alternative Methods 3.5 Acoustics 4 Big Data for Powertrains 4.1 The Future of Powertrain Concepts 4.2 The History of Simulation 4.3 Development Processes and Scenarios of Simulation with Big Data 5 Powertrain Development with Artificial Intelligence 5.1 Models of Artificial Intelligence 5.2 Overview: Levels of AI 5.3 Search Tree 5.4 Machine Learning 5.4.1 Supervised Learning 5.4.2 Unsupervised Learning 5.4.3 Reinforcement Learning 5.5 Artificial Neural Networks (ANN) 5.5.1 Activation Function 5.5.2 Feedforward Networks (ANN) and Recurrent Networks (RNN) 5.5.3 Training Procedure 5.5.4 Network Architecture and Performance 5.5.5 Fuzzy Logik 5.6 Deep Learning 5.6.1 Convolutional Neural Network (CNN) 5.6.2 Generative Adversarial Network (GAN) 5.7 Software Bibliography 3.2 State variables, fluid and energy flow of a container model 3.3 Flow function as a function of isentropic exponent and pressure ratio 3.4 Examples of EHD-lubricated contacts 3.5 Stribeck curve: Effect of the lubricant film on the friction coefficient of two materials 3.6 Coordinate system in the contact point of two bodies 3.7 Physical/chemical process sequences in a cylinder 3.8 Exhaust gas equilibrium composition for C8H18 fuel 3.9 Partial equilibrium using the example of the extended Zeldovich mechanism 3.10 Geometry of the crank mechanism 3.11 2D and 3D finite elements with linear, quadratic and cubic deformation 3.12 Finite element for a 2D surface with a large and a small mesh size 3.13 Force balance on a body 3.14 Deformation of a body 3.15 Shear deformation of a body 3.16 Transversal contraction of a body 3.17 Number of nodes per body edge by a displacement function 3.18 Displacement of all nodes of a body represented by the displacement vector 3.19 Time and frequency spectrum by superposition of harmonic signals 4.1 Emission limits according to the European exhaust standard 4.2 Trend of new passenger car registrations along powertrain concepts in Germany until 2040 4.3 Trend of the combustion engine as a proportion of all new car registrations 4.4 Trend of passenger car stock along driving concepts in Germany until 2040 4.5 History of conventional simulation tools for general applications 4.6 History of conventional simulation tools for the automotive industry 4.7 Development trend of all simulation software products 4.8 Development trend of emission limits since the introduction of emission standards 4.9 Specific CO2 and NOx fleet development trend since the introduction of emission standards 4.10 Specific CO2 fleet development trend since the introduction of CO2 limits 4.11 Characteristics of data (Big Data) 4.12 Prototype-oriented/test-based development process of powertrains up to the1970 s 4.13 Simulation-oriented development process of powertrains today 4.14 Historical change of development processes through simulation 4.15 Generating big data through real data or through simulation software 4.16 Trade-off between level of detail and model accuracy: AI simulation compared to conventional simulation 4.17 Serial tasks vs. parallel tasks 4.18 Speed advantage GPU vs. CPU 4.19 Gadgeting process: Rationalization and reassortment of all simulation tools to the strongest tool in its respective discipline, extension of the development landscape with AI gadgets 4.20 Gadgeting sub-concept: AI gadgets to support the calibration 4.21 Typical procedure of a model calibration and applying a model as a data generator 4.22 Method for big data generation and creation of AI gadgets 4.23 Development of a prototype production: Gadgeting sub-concept 4.24 Gadgeting full concept: AI gadgets as a result supplier 4.25 Development process of a prototype production: Gadgeting full concept 4.26 Submodels in the development/simulation of combustion engines 4.27 DoE approaches for the creation of an experimental space 4.28 Integration of AI into the V-development lifecycle for automotive applications 5.1 Singularity 5.2 Proportion of SME and LE that already use AI technologies 5.3 Proportion of SME and LE which work with external AI providers 5.4 Proportion of SME and LE that use AI technologies at least to a small extent today and probably in five years 5.5 Flowchart: Levels of artificial intelligence 5.6 Search tree structure 5.7 Search tree through heuristic downward strategy 5.8 Optimizing ECU application parameters for a lean concept engine using search tree 5.9 Systemic representation of a parallel-serial hybrid powertrain 5.10 Transient application of search tree: Optimization of a hybrid control strategy for the WLTC 5.11 Traditional programming vs. machine learning 5.12 Learning methods for machine learning 5.13 Classification and regression 5.14 Observer for the supervised Learning process 5.15 Observer combined with a physical model 5.16 Classification of thermokinetic data (Reynold number) 5.17 Engine model of a control unit 5.18 Regression of thermokinetic data (Reynold number) 5.19 Clustering and association of data for the Unsupervised Learning procedure 5.20 Interpreter for the Unsupervised Learning process 5.21 Normal distribution of measurement outliers (quantization error) of a data set 5.22 Data selection for two different quantization error thresholds 5.23 Fuel cell: Grouping of efficiency classes 5.24 Clustering to investigate a steel with nickel-chromium-molybdenum alloy after material failure 5.25 Agent for the Reinforcement Learning process 5.26 Static control circuit for the vehicle speed of a cycle 5.27 Adaptive control loop through Reinforcement Learning 5.28 Adaptive RDE speed control 5.29 Reinforcement Learning applied to chassis design 5.30 Functioning of today's computer architectures based on Neumann 5.31 Components of a neural network 5.32 Activation functions 5.33 Schematic representation of internal processes between two connected neurons 5.34 Feed-Forward ANN with 3 levels: Hinton diagram with pure feed-forward formation and additional direct connections between an input and an output 5.35 Feed-Forward ANN with self-aligned recurrent loops of the neurons and externally aligned recurrent loops 5.36 Feed forward ANN with recurrent loops only within the layers and completely linked ANN 5.37 Prediction of time-dependent variables: Definition of past time step and future time step 5.38 Merging sequential RNNs for processing time-based events 5.39 Functionality of an LSTM cell 5.40 Training and prediction of CO, NOx emissions of the RDE test procedure 5.41 Prediction of CO and NOx in RDE for different future times steps 5.42 Backward propagation training phase 1: Feed forward and derivation of individual neural outputs 5.43 Backward-Propagation training phase 2: Derivation of errors and gradients 5.44 Backward propagation training phase 3: Update of the weights based on the gradients 5.45 Neural network for counter-propagation training 5.46 Projection of data onto a 2D feature map (Self Organized Maps SOM) 5.47 Performance of an ANN as a function of data volume and network architecture 5.48 Sequential construction of a neural network according to a serial method and parallel method 5.49 Training of a serial and parallel network 5.50 Network dimension vs. efficiency 5.51 Number of hidden layers to be selected depending on data volume and number of input variables 5.52 Number of hidden units to be selected depending on the amount of data 5.53 Dropout of neurons 5.54 Subdivision of a data set into batches 5.55 Training/Test vs. epochs 5.56 Temperature sensation in a group of test persons 5.57 Combination of fuzzy logic and a neural network as a hybrid system for 1. ANN as data supplier and 2. fuzzy logic as data supplier 5.58 Weighting of different load shares of a driving cycle using fuzzy logic 5.59 Cascade structure of a CNN process 5.60 Conversion of an RGB pixel code by a convolution filter (convolution process) 5.61 Pooling method for data filtering 5.62 Geometries of three different fuel injectors 5.63 Optical measurements of injection characteristics for different geometries 5.64 Combustion optics: Application of fluorescence filters to enhance image features: Diffusion flame and soot formation 5.65 CNN to predict specific quantities (turbulence, mixture formation, emission, etc.) 5.66 Training phase of a flow through a plate with variable geometry 5.67 CNN for generating time-discrete outputs of flow results 5.68 Validation of the CNN model using flow-through plates with different incisions 5.69 Plate flow with central, circular incisions in comparison: CNN versus 3D-CFD 5.70 Plate flow with two circular incisions on the transverse axis in comparison: CNN versus 3D-CFD 5.71 Plate flow with asymmetrical incisions in comparison: CNN versus 3D-CFD 5.72 Conventional Sound Recognition: Sequential extraction of individual sound features 5.73 Sound Recognition: Sequential extraction of individual sound features by CNN 5.74 Sound recognition applied to the frequency spectrum of an engine to recognize turbocharger specific noises 5.75 Functioning of the discriminator and the generator at microscopic fracture structures of a metal material 5.76 Working process of a GAN 5.77 GAN for an extremely scalable generation of data 5.1 Milestones in artificial intelligence 5.2 Data processing brain versus computer 5.3 Open-source software and examples of some libraries for the application of AI