Complete standalone solution for sensor data: from learning to inferenceSolist-AI^™ Achieving Groundbreaking Embedded AI

When we think of AI, vision AI often comes to mind. It is used to recognize objects in images, identify faces, and inspect product images. More recently, generative language AI, which can produce text based on extensive training, has garnered significant attention. These types of AI are designed primarily to duplicate human intellectual activities, such as recognition and creation.
At the same time, other types of AI are being sought for fields such as industrial equipment. For example, AI that can detect subtle changes in sensor signals to predict failures or detect abnormalities. ROHM introduces Solist-AI^™, a groundbreaking AI solution optimized for such applications, capable of performing the entire process, from learning to inference, directly on embedded devices. By integrating a dedicated hardware AI accelerator within the MCU, Solist-AI^™ enables the implementation of embedded AI systems with virtually zero software overhead. This article explores ROHM’s Solist-AI^™ and its diverse applications.

Conventional AI vs Solist-AI^™

AI technology is advancing at a rapid pace, giving rise to various types of AI, including speech recognition AI that identifies words from human speech, vision AI that recognizes objects and their conditions in images, and language generation AI that creates text in response to questions. Today, AI has become a foundational infrastructure supporting activities across daily life and industry.
To function effectively, these AIs must undergo extensive training with large datasets to develop inference models for recognition and classification, as well as large-scale language models for text generation. This training is typically conducted on cloud computers capable of processing vast amounts of data at high speeds. When it comes to inference processing, several approaches are available.

●Cloud AI

This approach uses high-performance cloud computers to directly run inference models. Data generated at the edge endpoints is sent to the cloud for processing, which raises issues such as Internet communication costs, latency, and concerns about data privacy and security.

●Edge AI

In this method, the inference model is simplified and deployed on edge computers, allowing inference to be performed locally while reducing computational load.

●Endpoint AI

A more advanced approach simplifies the inference model even further to enable inference directly on endpoint devices, such as SoCs or MCUs. ROHM proposes Solist-AI^™, a next-generation AI solution that goes beyond conventional AI by performing learning at the endpoint, operating as a standalone system that eliminates the need for network or cloud connectivity (Fig. 1, Table 1).

Solist-AI^™ Features

In industrial equipment operating in factories and on-site, monitoring operating conditions and detecting abnormalities often requires the use of various sensors. Traditionally, this type of monitoring has been performed without AI, relying instead on simple circuits and software to detect signals that exceed predefined thresholds.
While conventional AI can technically be used for such monitoring, factors such as the effort required to train models on the cloud, increased costs and power consumption associated with edge-side processors, and the expenses of cloud services and internet communication, often make adoption impractical.
Solist-AI^™ offers a different approach by implementing the entire AI monitoring system directly on the device operating at the site. This allows Solist-AI^™ to not only detect abnormalities after they occur, but also identify potential issues from subtle changes in sensor signals. This capability is particularly useful for predictive maintenance.
With Solist-AI^™, learning occurs at the endpoint, allowing the system to adapt to the specific environment and unique characteristics of each device or piece of equipment. Moreover, further learning and retraining can be carried out on-site without relying on external infrastructure.

How Solist-AI^™ Works

Solist-AI^™ employs a three-layer neural network consisting of an input layer, a hidden (intermediate) layer, and output layer, performing learning and inference using input data such as sensor signals. What’s more, the network can compress input data, reconstruct it, and output the results at high speed using the same hardware circuitry.
During the learning phase, parameters like the intermediate layer and network weights are adjusted to align the output data with the input data. Initially, an untrained Solist-AI^™ shows a low level of agreement between input and output, but as training progresses, the consistency improves for the learned data.
In the inference phase, if an unknown input signal matches the learned data, the input and output will be consistent, but for unlearned data, the input and output will diverge significantly. By training the system with normal sensor signals, Solist-AI^™ can detect deviations from the normal values based on the magnitude of discrepancy.

32-bit MCU for Solist-AI^™

To implement Solist-AI^™, the MPU or MCU used in endpoint devices must be equipped with AI learning and inference capabilities. ROHM has begun mass production of the ML63Q2500 Group of 32-bit MCUs as the first products in the lineup.
The ML63Q2500 Group is powered by an Arm^Ⓡ Cortex^Ⓡ-M0+ core and features the AxICORE-ODL hardware accelerator that enables high-speed execution of Solist-AI^™ learning and inference processes (Table 2).
(*The AxICORE-ODL hardware accelerator: A functional block that improves the processing speed of AI functions by utilizing hardware circuits instead of software by CPUs.)
Designed for industrial applications, a comprehensive set of functions are built in, including a CAN FD controller, 3-phase motor control PWM, and 12-bit ADC x 12 channels.

A standout feature of Solist-AI^™ is that all AI computations, both during learning and inference, are handled entirely by the AxICORE-ODL hardware accelerator. This allows the CPU to remain available for processing non-AI programs, with the software load for AI operations being virtually zero.
AxICORE-ODL delivers millisecond-level response time, executing tasks in approx. 1/1000th the time required for software-based processing.
(*compared to ROHM’s conventional product)

ML63Q2500 Group Development Support Tools

The ML63Q2500 Group of standard Arm^Ⓡ core MCUs supports the use of general-purpose tools for software development. Additionally, ROHM offers its own Arm^Ⓡ integrated development environment. For Solist-AI^™ development, ROHM provides tools such as an AI simulator that runs on a PC along with a real-time viewer for visualizing internal MCU data.

There is no need for cloud-based tools or AI frameworks such as TensorFlow^Ⓡ commonly used in conventional AI development.

To lower the barrier to AI adoption for customers, ROHM is developing an ecosystem in collaboration with partner companies. By providing a wide range of easily deployable solutions, ROHM allows customers to achieve immediate implementation.

• Arm^Ⓡ and Cortex^Ⓡ are registered trademarks of Arm Limited (or its subsidiaries) in the US and other countries.
• TensorFlow^Ⓡ is a trademark or registered trademark of Google LLC.

• Solist-AI^™ is a trademark or registered trademark of ROHM Co., Ltd.
Solist-AI^™ is the brand name for ROHM's on-device AI solution designed for edge computing applications. Drawing inspiration from the musical term "Solist," which signifies solo performance, this innovative technology enables real-time learning and inference directly on edge devices without relying on cloud servers. Powered by ROHM’s proprietary on-device learning AI technology (or chip), Solist-AI^™ is characterized by its compact design and low power consumption, contributing to the expansion of sustainable AI innovation.