AI MCUs features standalone learning and inference

ROHM has released the industry-first AI MCUs capable of standalone learning and inference, specifically developed to enable predictive maintenance.

The new ML63Q253x-NNNxx and ML63Q255x-NNNxx series employ real-time sensor data to detect anomalies, predict failures, and anticipate equipment degradation across a wide spectrum of industrial applications.

As industrial machinery becomes more intelligent and efficient, the ability to perform early fault detection and predictive maintenance provides significant operational and economic advantages.

In an exclusive interview with Embedded.com, Kenichi Morioka, technical marketing manager at ROHM Semiconductor USA, emphasised that these devices are the first in the industry capable of performing AI learning and inference entirely on the device, eradicating the necessity for network connectivity.

"Solist-AI digitises and learns the normal operating state of target equipment on-site using sensor input, then detects deviations from this baseline as an 'anomaly score' – a numerical indicator used to predict potential failures or performance degradation," explained Morioka.

By conducting learning directly at the deployment site, the system can accurately capture the specific characteristics of each device, including environmental factors such as noise, vibration, and humidity, and individual unit variations, enabling high-accuracy failure prediction.

Morioka added, "This technology simplifies deployment through on-site learning without relying on the cloud. Operating without a network connection ensures faster response times, minimises information security risks, and reduces communication costs."

Solist-AI captures and learns normal operating parameters, such as motor current, temperature, and rotational speed, employing data collected from sensors.

"If abnormalities like bearing wear or load misalignment occur, they manifest as subtle changes current and vibration patterns. Solist-AI identifies these deviations and raises the anomaly score accordingly," Morioka explained.

In a typical implementation, this elevated anomaly score is transmitted to a host MCU, which interprets this signal as an indication that maintenance is required.

"Equipment failures can lead directly to line stoppages and increased maintenance costs," noted Morioka.

Proactively addressing potential issues before a complete failure occurs is essential to minimising downtime and reducing overall maintenance expenses, making predictive maintenance a critical and widely adopted practice.

Morioka further commented, "In some countries, regions, or installation sites, it can be challenging to secure inspection personnel or conduct frequent checks. We believe this solution offers a practical and effective alternative in such cases."

Solist-AI includes a compact three-layer neural network algorithm that allows learning and inference directly at the endpoint, such as motors, without requiring cloud connectivity, host computers, or network networks.

"To support smooth evaluation and implementation, we offer development tools such as Solist-AI Sim, a simulator for verifying AI behaviour, and Solist-AI Scope, a real-time visualisation tool that illustrates AI performance," Morioka explained.

Solist-AI Sim, in particular, enables users to assess the suitability and effectiveness of Solist-AI for their specific application on a PC before deploying the hardware.

Solist-AI performs learning and inference directly on the MCU, eliminating the necessity for cloud services or large-scale data transmission.

"Anomalies can be easily communicated to a host microcontroller via standard serial interfaces, making it easy to retrofit Solist-AI into existing systems," Morioka noted.

Conventional AI models often rely on network access and powerful processors, introducing challenges like latency, higher costs, and cybersecurity risks.

In contrast, Solist-AI attains processing speeds up to 1,000 times faster than conventional software-based approaches while consuming only a few tens of milliwatts.

"This FLASH MCU is built around an Arm Cortex-M0+ core and comes equipped with a range of peripherals, including multiple serial interfaces, an A/D converter, CAN controller, multifunction timers (i.e. PWM and capture), and an analogue comparator," Morioka emphasised.

Beyond these standard features, integrating Solist-AI allows on-site learning and inference. Unlike conventional endpoint AI solutions that rely on cloud-based training and local CPUs or GPUs for inference, the company's approach integrates the AxlCORE-ODL AI accelerator to handle both processes efficiently on-device.

"As no communication with the cloud or external host is required for inference, and given its high processing speed, the system can detect anomalies and deviations from normal behaviour in real time," Morioka added.

Solist-AI is a fully standalone AI capable of performing learning and inference independently.

"Its versatility allows it to be applied in ways we hadn't initially envisioned," noted Morioka. "We plan to expand the product lineup to meet emerging needs."