Proceedings

Designing deep learning, computer vision, and signal processing applications and deploying them to FPGAs, GPUs, and CPU platforms like Xilinx Zynq™ or NVIDIA^® Jetson or ARM^® processors is challenging because of resource constraints inherent in embedded devices. This talk walks you through a MATLAB^® based deployment workflow that generates C/C++ or CUDA^® or VHDL code.

For system designers looking to integrate deep learning into their FPGA-based applications, the talk helps teach the challenges and considerations for deploying to FPGA hardware. We will briefly show how to explore and prototype trained networks on FPGAs using prebuilt bistreams from MATLAB. You can further customize your network to meet your performance requirments and hardware resource usage, generate HDL, and integrate it into an FPGA-based edge inference system.

While many organizations get excited about adopting machine learning techniques, success does not come easy. Come to this talk to learn about applications where machine learning generates considerable ROI, including fleet data analysis, energy forecasting, and smart manufacturing.  We will also demonstrate how engineers are integrating machine learning techniques with their controls and signal processing workflows to improve system performance.

Throughout the presentation we will highlight new features in MATLAB^® that accelerate deploying machine learning. This includes applying automation techniques to feature selection, model selection, and hyperparameter optimization (AutoML).  We will also cover new ways for integrating machine learning models with production workflows such as updating deployed models and C/C++ code generation.

Learn how your peers have applied machine learning, and to get inspiration for how machine learning could be applied to your own work.

Predictive maintenance reduces operational costs for organizations running and manufacturing expensive equipment, by predicting failures from sensor data. However, identifying and extracting useful information from sensor data is a process that often requires multiple iterations as well as a deep understanding of the machine and its operating conditions.

In this talk, you will learn how MATLAB^® and Predictive Maintenance Toolbox™ combine machine learning with traditional model-based and signal processing techniques to create hybrid approaches for predicting and isolating failures. You will also see built-in apps for extracting, visualizing, and ranking features from sensor data without writing any code. These features can then be used as condition indicators for fault classification and remaining useful life (RUL) algorithms.

Software algorithms play a critical role in battery management systems (BMS) to ensure maximum performance, safe operation, and optimal life of battery pack under diverse operating and environmental conditions. Developing and testing these algorithms requires expertise in multiple domains and achieving functional safety certification can be very confusing and lengthy process.

In this talk, you will learn how to:

• Design and test BMS algorithms such as state of charge estimation, cell balancing, contactor management, and current/power limit calculation
• Generate production quality C/C++ code and target embedded processors
• Measure design complexity and perform systematic unit testing
• Prove that your design meets requirements and automatically generate tests
• Perform hardware-In-th-loop testing using Speedgoat real-time hardware
• Produce reports and artifacts and certify to functional safety standards

Grid-tied inverters connect renewable energy sources to an electric utility grid. This session will show you how to model, simulate, and implement a controller for a grid-tied solar inverter using Simulink® and Simscape Electrical™.  The demo will use a photovoltaic (PV) inverter to show you how to:

• Build a model to simulate the PV panel, grid load, and grid-tied inverter
• Tune the inverter controller to adjust the DC bus voltage for varying loads
• Develop a maximum power point tracking (MPPT) algorithm to keep the PV system operating at the peak power point under changing solar irradiance, temperature, and load
• Implement plant level control such as fault-ride through, curtailment, and power factor correction
• Generate C code from the controller model and implement it on a Texas Instruments C2000™ microcontroller
• Test the microcontroller code against a real-time hardware-in-the-loop (HIL) simulation of the PV system and inverter using a Speedgoat real-time system

Brushless motors are everywhere. These motors rely on field-oriented control to regulate currents to motor windings. In this session, we will look at some of the challenges and solutions for developing field-oriented control algorithms to achieve the 10-20KHz switching frequency required for efficient and accurate motor operation.

Highlights:

• Approaches for automatically generating fast compact floating-point and fixed-point code from models
• Methods for developing sensorless field-oriented control algorithms
• Approaches for tuning speed and current loop gains
• Techniques to model and parameterize electrical machines for simulation work

Systems engineering is a challenging problem, and often the tools used to tackle these challenges do not connect well to the other tools used throughout the design process. MathWorks systems engineering tools combine with MATLAB^® and Simulink^® to create a unified modeling environment, enabling the use of a single platform throughout systems engineering, design, implementation, and verification processes.

In this talk, we present a workflow for systems engineering and architectural modeling with a tight connection to Model-Based Design.

Highlights:

• Building a bridge between early architecture work and downstream design
• Creating architecture models and extending the language through stereotypes and profiles
• Analyzing architectures
• Moving to design and implementation

Tractor semitrailers are widely used in the transportation industry for moving heavy-duty goods from one place to another. They operate well over an eight hour range per day, so analysing ride comfort characteristics is paramount for minimizing fatigue and stress for the driver during uptime of the vehicle. ISO 2631-1 serves as a guiding manual for the measurement and defines acceptable criteria for ride comfort characteristics of the vehicle. Tractor- semitrailers pose a major challenge due to mating of the load body on the tractor frame via the fifth wheel coupling mechanism. This independency causes the load body to act as a source of vibration and transmits the force onto the driver seat via the chassis. The location of the fifth wheel coupling is critical for ride comfort characteristics of the vehicle. Furthermore, accurate establishment of ride comfort parameters in the virtual environment saves cost and time developing prototypes and iterating on the physical vehicle. This project thus focuses on the establishing ride comfort parameters on the seat level of a 4x2 heavy tractor- semitrailer. Thus, a detailed 7-degree-of-freedom mathematical model is developed to analyze the characteristics of the vibrations in a virtual environment. The mathematical model is developed using MATLAB^® and Simulink^® R2019b. Consequently experimentation is carried out on a physical vehicle to obtain acceleration- time domain data by mounting accelerometers on various locations. The vehicle is run on A and B class road profiles to account for random vibrations. The same road vibrations are used as an input to the mathematical model. Similarly, a simulation is done on MATLAB and Simulink and a correlation is established between simulated and actual data using time domain and frequency domain analysis. Furthermore, the data is fitted on the eight-hour curve as per ISO 2631-1. Thus a versatile, parameterized mathematical model for ride comfort analysis of a tractor- semitrailer is developed.

As simulation becomes increasingly important across an organization, engineering teams need to share their Simulink^® simulations with adjacent groups, suppliers, and clients.

In this presentation, you will see how to share Simulink simulations as standalone executables and browser-based web apps that can be accessed  with unique URLs. In addition, you will see how to deploy your simulations on MATLAB Production Server™ as APIs that can be called from enterprise applications.

Reinforcement learning allows you to solve control problems using deep learning but without using labeled data. Instead, learning occurs through multiple simulations of the system of interest. This simulation data is used to train a policy represented by a deep neural network that would then replace a traditional controller or decision-making system.

In this session, you will learn how to apply reinforcement learning using MATLAB^® and Simulink^® products, including how to set up environment models, define the policy structure, and scale training through parallel computing to improve performance.

Lane keeping assist (LKA) and lane following assist (LFA) systems are an important feature of advanced driver assistance systems (ADAS) and automated driving (AD). They automatically take action to keep the vehicle in its lane and follow cars at a safe distance. Variations in road and driving conditions must be considered when designing reliable and robust LKA and LFA systems.

Real-time simulation and testing can help verify these systems with simulated driving. We will discuss how to simulate road and other conditions using a photorealistic simulation environment leveraging the Unreal Engine from within Simulink^®. Using a multicore Speedgoat system, we simulate the vehicle dynamics and powertrain as well as lane keeping and following with model predictive controls. The vision-based lane detection algorithm is implemented on a Simulink programmable FPGA I/O module from Speedgoat, making it possible to process video data from the Unreal Engine.

In order for autonomous systems to move within their environment, engineers need to design, simulate, test, and deploy algorithms that perceive the environment, keep track of moving objects, and plan a course of movement for the system itself.  This workflow is critical for a wide range of systems including self-driving cars, warehouse robots, and unmanned aerial vehicles (UAVs).  In this talk, you will learn how to use MATLAB^® and Simulink^® to develop perception, sensor fusion, localization, multi-object tracking, and motion planning algorithms.  Some of the topics that will be covered include:

• Perception algorithm design using deep learning
• Fusing sensor data (cameras, lidar, and radar) to maintain situational awareness
• Mapping the environment and localizing the vehicle using SLAM algorithms
• Path planning with obstacle avoidance
• Path following and control design

Learn about new capabilities in the MATLAB^® and Simulink^® product families to support your research, design, and development workflows. This talk highlights features for deep learning, wireless communications, automated driving, and other application areas. You will see new tools for preprocessing and analyzing data; developing motor control algorithms; creating interactive apps; packaging and sharing simulations; and modeling, simulating, and verifying designs.

Organizations with digital transformation initiatives are making the transition from visionary ambitions to practical projects.  These organizations have defined their high-level digital transformation objectives, and are now looking to their engineers and scientists to achieve them by learning new technologies, collaborating with unfamiliar groups, and proposing new products and services.

To meet this challenge, technical organizations must master how to systematically use data and models, not only during the research and development stages, but also across groups throughout the lifecycle of the offering.  An effective digital transformation plan needs to consider changes in people’s skills, processes, and technology. 

Join us as Jim Tung describes this pragmatic approach to digital transformation and demonstrates how engineering and scientific teams are leveraging data and models to achieve their transformative objectives.

Owing to their harsh operating conditions, oil-filled transformers tend to degrade over time and leak oil, which results in failure or reduced life. The main indicators of transformer health are oil level and top oil temperature. Further, being highly optimized for cost, these transformers have very minimal instrumentation. Our solution proposes a scheme to virtually sense the oil level using temperature sensors fitted on the transformer housing. A physics-based model of the transformer enables us to relate the temperature measurements to the amount of oil present in it in real-time. This non-invasive solution can be retrofitted on any ONAN transformer and can enable real-time condition monitoring of the asset.

The next step in implementation of our solution was field deployment. We exploited the features of the MATLAB Production Server™ to deploy our algorithms to the cloud seamlessly. Prototype IoT devices deployed on individual transformers stream data to the cloud where our algorithm generates signals for predictive maintenance of the respective assets.

We present a predictive manufacturing solution to reduce manufacturing scrap in their optical manufacturing process. The ingredients here are MATLAB^®, Machine Learning and a new piece of automation tech—Robotic Process Automation (software bots). To understand how this ballet of tech creates tangible impact on the company's bottom line, here's a primer on optical fibre manufacturing. STL uses a process called RIC (Rod in Cylinder) to manufacture an optical fibre product from their portfolio. Basically, you put a no. of glass rods (higher refractive index) in a big glass cylinder (lower refractive index) and fuse them together to get a larger glass cylinder called preform. After that you load this preform on top a multi-storey building called draw tower and melt the bottom part in a furnace. What happens is the glass material starts dripping through a very controlled process, you get a very thin strand of optical fibre.

Our aim is to improve the design margin by providing the electrical signature of a fault on various subsystems dependent in a direct or indirect way. The framework allows us to run iterations on possible combinations of faults at subsystem levels as well as at wire fault levels and use the recorded data for impact analysis of subsystem failures on the entire system. While it increases the number of cases in the failure mode analysis, it results in a significant reduction of the cost and time required for testing. The inputs are taken from an optimized orthogonal matrix generated based on the probability of occurrence of a fault out of the manifold combinations of fault injection possible in the framework. The effect is generated by a high-fidelity comprehensive electric vehicle model that facilitates the injection of these faults and provides data for root cause identification of observed failures, thus making way for future prognostics.

Electric vehicles are powered by lithium-ion batteries which require an effective but low power consuming battery thermal management system (BTMS). In this presentation we’ll learn about how 1D simulation has been used in MEML to optimise the BTMS. The ease of 1D modelling allows it to be used during concept development and in parameter tuning to choose the best components at its best operating point. The model explained here consists of a driver model, vehicle model, equivalent circuit model, battery box model, and refrigeration cycle model. After validating the model with test data, thermal algorithms have been tested for optimizing energy requirements. In this system model, the battery, cooling circuit, and refrigeration circuit were effectively implemented using Simscape^®. The vehicle, driver, and equivalent circuit model were implemented on the Simulink^® platform.

Today, several processes in a typical industrial plant scenario are modelled and simulated in Simulink^® to evaluate control and process timings and tuning. However, to be able to deploy and tune with the plant dynamics, engineering and commissioning tools that integrate with the process communication are used. This restricts the design tools such as Simulink to perform parameter sweeps and optimization and in general breaks the model-based development philosophy.

With MathWorks support, we have achieved code generation for industrial asset targets for Siemens' SINAMICS drives. Users are now able to generate model blocks and deploy them to commissioning tools and even get feedback on the performance of the models helpful for parameterization. The latest code-generation techniques using coder descriptors help greater customization and maintainability of the target files. This talk explains the nuances and experiences of using and integrating MATLAB^® and Simulink with industrial tools and assets.

Radar's advantage and application in mobility and ADAS are limited to object detection and position estimation. Very limited developments concentrate on object classification based only on radar. This is due to the limitation of a less dense point cloud from radar. At PathPartner we have worked on classification of vulnerable road users based only on radar Data. MATLAB^® and Statistics and Machine Learning Toolbox™ have helped us analyze and decide on the most efficient and suitable algorithm for this application. We have developed this algorithm with the toolbox and have implemented it in an embedded platform and verified it in actual test scenarios.

In this presentation we would like to present the advantages and challenges of using the deep reinforcement learning for closed loop powertrain control algorithms. The powertrain state feedback control functions have to consider a huge variety of environmental conditions. The way to develop and customize this new generation of control functions will change our traditional system and software process. The use of artificial intelligence and in particular of deep reinforcement learning techniques will speed up the development and improve the performance and contribute to higher efficiency.

Applying verification and validation techniques early in the development process enables you to find design errors before they can derail your project. Most system design errors are introduced in the original specification but are not found until the test phase. When engineering teams use models to perform virtual requirement based testing early in a project, they eliminate problems and reduce development time.

In this master class, you will learn how you can apply early verification and validation activities at every stage of the development process to ensure that your design meets the functional requirements and is free from any design errors.

Highlights include:

• Detecting design errors in models using formal verification methods
• Systematic simulation testing of design by using an automation framework
• Automatically generating reusable tests that satisfy model and code coverage
• Formally validate safety-critical properties and verify design robustness.