Day 1 | Free | Accelerated AI for IoT made efficient – The VEDLIoT project

The VEDLIoT project targets the development of energy-efficient Deep Learning for distributed AIoT applications. A holistic approach is used to optimize algorithms while also dealing with safety and security challenges. The approach is based on a modular and scalable cognitive IoT hardware platform. VEDLIoT offers a complete design flow for Next-Generation IoT devices required for collaboratively solving complex Deep Learning applications across distributed systems. The methods are tested on various use cases ranging from Smart Homes to Automotive and Industrial IoT appliances.

Honey.AI is a novel device utilizing robotics, computer vision, and deep learning to provide real-time, cost-effective quality assessments for honey. Traditionally, pollen analysis, essential for quality control and fraud detection, is manual, time-consuming, and error-prone. Honey.AI aims to transform this process, by offering onsite assessments. The goal of the VEDLIoT experiment is to enhance Honey.AI by incorporating edge AI computing within the robot and boosting the performance and development agility of the AI modules using EmbeDL and Kenning.

The MushR project leverages digital technologies and intelligent automation to transform gourmet mushroom cultivation, enhancing productivity, efficiency, and quality. Controlled environment tents with smart sensors maintain optimal growth conditions, while a Mask R-CNN model integrated image capture system enables real-time maturity detection. A semi-automated harvesting system uses VEDLIoT near-edge computing for precise harvesting. The system’s modularity allows for industry-level scaling, drastically reducing manual labour and improving yields and quality. The project highlights the potential of digitization in agriculture, promoting advancements in digital farming.

Control algorithms for electrical drives are crucial for efficient operation in diverse applications. Traditional control algorithms can be complex, inefficient, and require expert implementation. Machine learning approaches, like reinforcement learning, offer adaptive, optimal control but incur high computational costs. The VEDLIoT project offers a solution by developing an IoT learning setup that ensures operational safety with FPGA resources for time-critical decisions, while also enabling asynchronous remote learning on an edge-computing workstation.

DUNE’s Real-Time Localization System (RTLS) harnesses distributed computing capabilities across far-edge, edge, and cloud levels for efficient real-time asset tracking applications. Outperforming traditional methods, DUNE’s deep learning model efficiently estimates emitted Bluetooth Low Energy signals’ direction on VEDLIoT hardware at the far-edge. It achieves sub-meter accuracy by combining estimates from far-edge nodes, preparing to incorporate additional wireless technologies. DUNE RTLS has been implemented in both pre-production and production environments, demonstrating practical application at the Universitat Oberta de Catalunya’s R&I hub.

BEAM-IDL addresses the increased demand for Aluminium in Electric Vehicle manufacturing with optimized Deep Learning algorithms for laser welding process image recognition. This innovative approach improves laser-material interaction, enhancing process control and optimization. Experimental tests were conducted using a robot cell and an industrial optic, with results improving algorithm performance. The integrated solution, merging AI with monitoring and control systems, offers competitive advantages against lower-value manufacturing alternatives, enhancing both EXOM’s products and LORTEK’s services.

AI-RIDE is a groundbreaking project that employs an accelerated, online, and embedded AI framework for motorcycle rider training, particularly in Practical Driving Courses and Driving Licence Exam verification tools. The platform, driven by VEDLIoT hardware and middleware, employs a disaggregated IoT and sensor-based network architecture for enhanced driving learning techniques. Utilizing distributed algorithms, high-speed networking, and embedded AI at the edge, AI-RIDE performs innovative functions using video and image processing to fuse data from various sources, enhancing the driving test session with optimal vision targets.

For the management of chronic diseases continuous remote monitoring and fast accurate alerts are of high importance. In this context, wearable innovations with on-device computation capabilities can be extremely valuable. The overall project goal of FLEDGED was to evaluate the next-generation wearables architecture for deep learning deployments in healthcare. During the project, the feasibility of improving wearable systems in 3 main VEDLIoT technology areas, spanning from infrastructure & tools in the centre to the far edge of the system, was assessed.

The AccBD project tackled the challenges of data analysis in medical research with high dimensionality and skewness, using the VEDLIoT hardware platform to accelerate a feature selection workflow. The results included a 13.9X acceleration and a 72% energy saving. The project also implemented the Yeo-Johnson transformation, integral for data preprocessing, as an optimized C-library and an FPGA module, offering substantial runtime and energy reductions. The results highlight the potential of optimized parallelization and alternative hardware for efficiency in medical research.

Edge4iwelli is a VEDLIoT Open Call project focused on integrating edge computing in smart homes for health and well-being care. It enhanced the iwelli IoT care package with AI modules and edge computing, introducing new tools like the health smart mirror. The project aims to enhance the adoption of edge computing in smart homes and improve healthcare decision support and remote patient management.

The FLAIR project achieved four key goals: 1) it added Federated Learning functionality to VEDLIoT solutions, 2) it established a realistic test environment within a 5G network, 3) it quantified various FL challenges through diverse real-world scenarios involving IoT devices with varying computational capabilities and local data distributions, and 4) it devised intelligent FL solutions for participant selection and resource allocation. This presentation will delve into the project’s undertakings and outline its future directions.

This intro serves as an intro to the technical deep dive of VEDLIoT, exploring the various facets of the project. The showcased applications thus far are all based on the underlying VEDLIoT technology, which covers an extensive range of conceptual components such as requirements engineering or architectural frameworks. Methods for security and safety considerations related to AIoT systems are explained. In addition, the cognitive IoT Hardware Platform is detailed, also covering hardware accelerators and Co-Design aspects.

The architectural framework for VEDLIoT enables smooth design and integration of both traditional and deep learning components, considering quality concerns early in the design process. Organized by clusters of concerns and levels of abstraction, it forms a 2D grid guiding VEDLIoT system design. This structure simplifies complexity and enhances traceability, encompassing all aspects from logical behaviour to privacy and energy. Supporting a middle-out engineering approach, the framework combines top-down design with the integration of specific lower-level components.

VEDLIoT uses hardware and system-level tools to bolster the dependability and security of edge applications. Trusted execution environments are used, combining hardware features and dependability techniques. There’s an emphasis on end-to-end trust, secure code execution, and communication using technologies like enclaves and open-source WebAssembly runtime. ARM SoCs are employed with TrustZone and the open-source operating system OP-TEE, implementing remote attestation for WebAssembly code and offering enhanced security.

VEDLIoT’s hardware development concentrates on expanding and enhancing existing platforms like RECS|Box and t.RECS, with the creation of uRECS, focuses on compact, cost-effective, and energy-efficient designs for AI and ML applications. A broad array of hardware accelerators for deep learning is being explored. Performance evaluation of these accelerators using DL models like ResNet50, MobileNetV3, and Yolo is pivotal for integrating the best-suited DL accelerators into the RECS platform, aligning it with the use cases.

VEDLIoT applications aim to address high energy efficiency, security, and safety requirements for IoT and edge computing platforms. Specific use cases include the Pedestrian Automatic Emergency Breaking system in the automotive sector, the Motor Condition Classification and Arc Detection for industrial IoT, and AI-based smart home interactions. The focus is on optimizing energy efficiency, deep learning technology deployment, and ensuring high security while testing the concepts through challenging use cases across different industries.

IntellIoT promotes the advancement of humanized IoT and AI devices and systems. It aims to stimulate a competitive ecosystem in healthcare, agriculture, and manufacturing. Leveraging technologies such as 5G, cybersecurity, distributed technology, AR, and tactile internet, it prioritizes end-user trust, security, and privacy.

IntellIoT enhances safety in the agricultural sector by incorporating human-in-the-loop intelligence in IoT environments for semi-autonomous agricultural vehicles and drones. Humans intervene remotely in uncertain situations, thus refining the AI models. This intervention can be performed by a farmer or a remote service provider beyond private 5G cells. VR technologies offer a 360° live stream from the vehicle, while IoT/edge and networking infrastructures ensure secure, reliable operation.

Meeting of the VEDLIoT advisory board, composed of independent, board-level experts, providing external feedback on strategic decisions, reflecting the perspectives of different stakeholder groups. Confirmed participants at proposal submission include representatives from ARM Ltd, Ericsson AB, Volvo AB, Honda Research Institute Europe, Wago Kontakttechnik GmbH, dSPACE GmbH, Miele & Cie. KG, and bill-X GmbH.