[Editor's Note] Recently, on-device AI-based products are being built up rapidly in various fields. In this situation, it is time for hardware-centered companies to utilize solutions that can easily and quickly achieve AI function implementation, shorten product development periods, and support preemptive market entry. On the 15th, the 2024 Start On-Device AI (2024 SODA) conference was held by e4ds news, and insights were shared on solutions for lightweighting and optimizing AI models for installation on low-power and low-spec boards.

We spoke with Hyunsoo Moon, Manager of STMicroelectronics, who presented on Edge AI at SODA 2024, about ST's Edge AI technology and development cases.

■ Introduction to ST’s Edge AI Technology and Solutions

ST provides hardware and software to process machine learning algorithms, including deep learning, even on small embedded platforms with small memory size and computational performance, such as MCU and MPU products.

STM32Cube.AI allows pre-trained deep learning models created in general-purpose frameworks such as TensorFlow Lite on a host PC to be converted to C-code-based models and applied to STM32 products.

NanoEdge.AI Studio provides an AutoML pipeline that automatically generates code that can be applied to STM32 products, including data preprocessing, machine learning algorithm recommendation and learning, as well as the input dataset required for the machine learning model you want to create, even if you have no experience or knowledge of machine learning.

STM32Cube.AI and NanoEdge AI Studio enable STM32 customers to quickly and easily apply machine learning algorithms to STM32 products.

■ What are the key market/industry requirements and what are the differentiating and unique features compared to competitors’ products (or existing products)?
br /> In the case of hardware that constitutes edge AI, there are many manufacturers, and the types of products and solutions provided are very diverse. The key to edge AI is to minimize the loss of pre-trained machine learning models and to configure hardware that can accelerate the computation of machine learning models at the edge.

ST's STM32 family consists of a variety of lineups based on low power, enabling edge computing with low power consumption. In particular, the software tools of STM32Cube.AI and NanoEdge AI Studio allow users to easily and quickly apply machine learning models to STM32 products, and various options are also provided for weight reduction to minimize machine learning model loss.

■ Introduction to major application applications and representative use cases encountered in industrial sites and daily life

Edge AI is being used in a variety of fields, from small things like wearables, home appliances, and robots to large things like industrial equipment, smart homes, smart buildings, and smart cities.

A representative example is predictive maintenance, which is implemented by adding modules to the motor. The operating principle is that the vibration of the motor is detected by a MEMS sensor in the module, and even minute abnormal vibrations that cannot be detected by humans are detected.

There is a system that uses vibration data from such minute motors to notify maintenance before failure. Aftermarket edge AI devices that simply add edge AI modules to existing devices like this have low investment costs and very low power consumption.

Common home appliances include drum washing machines. Applying edge AI can increase energy efficiency by up to 40%. Drum washing machines are difficult to measure the actual weight of laundry due to the rotating drum, but this AI system measures the current when the motor first starts to operate to predict the weight of the laundry and determines the detergent, water, and washing time according to that weight to optimize energy efficiency. It is more efficient, reduces energy consumption, and can also save on maintenance costs.

In the security field, CCTV can be used to independently recognize video images on the device itself without using the cloud with the STM32N6 MCU. It automatically recognizes cars, trucks, buses, motorcycles, and pedestrians and displays them by color.

It operates at 5M pixel HD quality 18 FPS, uses only 0.5 watts of power, and costs less than 1 cent per day. If a CCTV system with AI functions is introduced to a smart farm, it will automatically recognize the appearance of wild animals in the farm and notify the farm owner.
■ Explain the future development direction and target market of the solution

Edge AI is driving revolutionary changes across a wide range of industries, and can significantly improve user experiences in many aspects, including real-time data processing, improved maintenance, and enhanced safety. ST is continuously updating its MCU functions to target the on-device AI market with AI functions.

In addition, ST plans to focus on the on-device AI market with new products that enhance security performance. Security standards will be strengthened in the US and Europe starting next year, and the low-power, high-performance MCU product line released this time has improved security features as well as power efficiency and performance. It is expected to expand its application to various high-performance AI fields as well as healthcare, door locks, and smart homes.

■ ST’s on-device AI ecosystem/collaboration status

Edge AI has many advantages, but AI developers must carry out various developments from start to finish. This edge AI development requires various hardware and software tools, and this can be solved through the ecosystem.

ST operates a very broad and comprehensive ecosystem, supporting boards, software, drivers, libraries, etc. Since the customer knows best about application development, ST plays a role in providing tools to help the customer develop applications well.

This ecosystem enables developers to easily develop edge solutions efficiently and reduce development time in a variety of environments.

■ What are your thoughts on participating in the event and the content presented at SODA 2024?

Participating in the 2024 SODA event once again confirmed the limitless growth potential of edge AI. It was even more meaningful as, from the perspective of a manufacturer that manufactures hardware for edge AI platforms, it was a valuable opportunity to directly introduce to customers the strengths of rapidly developing edge AI technology when combined with STM32 products.

ST will always provide the best support possible to help customers' products connect with the world faster and smarter through machine learning models.