The AI Revolution Starting in the Field
Over the past few years, artificial intelligence (AI) has transformed the world within massive data centers and the cloud. However, its full potential has not been realized right where data is generated—at the edges of factory equipment and machinery. This is because it has encountered the realities of network latency, security issues, and massive costs. Now, this paradigm is beginning to shift fundamentally. AI is leaving the "brain" of the cloud and seeping into small chips on the ground.

What if a small chip attached to a motor running ceaselessly in the middle of a factory could learn the machine's minute vibration patterns on its own and warn you in advance that "a bearing problem is likely to occur soon"? The Japanese semiconductor company ROHM has turned this very imagination into reality. The 'On-Device Learning MCU' developed by ROHM is an innovative semiconductor that completes both learning and inference on the device itself without a cloud connection.

This demonstrates that a new era is dawning where AI learns and grows directly in the field, going beyond the existing 'Edge AI' that simply runs models trained in the cloud.

Why did AI have to leave the cloud?

Despite their powerful performance, existing cloud-based AI models have revealed clear limitations in areas such as factory automation and real-time equipment control. Understanding these limitations is the first step in grasping the innovativeness of ROHM's technology. Existing models had several fundamental problems that prevented them from meeting the core requirements of industrial sites.

• Network Dependence and Latency: Cloud AI operates on a structure where data collected in the field is transmitted to a central server via a network, and the analyzed results are received back. The communication latency occurring during this process is critical in precision control processes where even a margin of error of a few milliseconds is unacceptable. In situations where minute abnormal vibrations in a motor must be detected and responded to immediately, network delays can lead to the shutdown of the production line or equipment damage.
• Security Risks: Transmitting sensitive information containing factory operational data and production know-how to the cloud via external networks implies exposure to serious security threats. Data breaches are a critical issue that can undermine a company's core competitiveness.
• Cost and Installation Issues: Running high-performance AI models requires powerful CPUs and large-scale network infrastructure, resulting in massive initial investment and ongoing operational costs. In particular, the challenge of 'retrofitting'—adding modern features to vast amounts of existing non-network equipment—is the biggest barrier to the widespread adoption of industrial AI and a problem that most cloud-based solutions have failed to overcome.

Ultimately, industrial sites are seeking new AI solutions that do not rely on the cloud, operate reliably even when network connections are down, and can be easily applied to existing equipment at a low cost. Why was the new approach known as 'On-Device AI' an inevitable choice as an alternative to overcome these limitations?

Self-learning and decision-making semiconductors

ROHM redefined the role of technology to address the previously raised industrial challenges. Their vision was to go beyond simply embedding AI into a small chip, enabling the chip itself to possess a complete intelligence. This stemmed from a clear development philosophy of "completing learning and inference on the device itself without going through the cloud."

It is precisely at this point that ROHM deliberately takes a more challenging path, deviating from industry conventions. While most 'edge AI' solutions merely execute pre-compiled models like remote limbs of a cloud brain, ROHM's philosophy is to empower the chip itself with the autonomy of learning. This approach directly overcomes the limitations of existing endpoint AI, which struggles to flexibly respond to subtle differences in installation environments or individual devices, even within the same model of equipment.

ROHM solved this problem through 'on-device learning.' Their AI MCUs learn the normal state of the equipment directly in the field and can perform self-relearning and additional learning in response to changes occurring over time. As a result, each piece of equipment possesses a customized AI model optimized for its own unique state. Since no cloud pre-training is required, it enables a truly standalone AI that operates perfectly even in environments without a network.

The core value this technology provides to customers is clear. It reduces maintenance costs by eliminating the need for network infrastructure setup, offers robust security free from the risk of data leakage, and ensures stable system operation by minimizing the risk of production line downtime through anomaly detection optimized for each piece of equipment. ROHM’s solution is to elevate semiconductors from simple computing devices into "intelligent entities that learn and make decisions on their own." So, how has ROHM technically realized this vision?

The operating principles of Solist-AI™ and AxlCORE-ODL

What made ROHM's vision a reality is the exquisite combination of software and hardware. At the core are the proprietary AI algorithm Solist-AI™ and the AI accelerator AxlCORE-ODL, which processes it at ultra-high speed. Understanding the operating principles of these two technologies allows for a clearer grasp of ROHM's innovation.

• Concept of 'On-Device Learning': Conventional edge AI was limited to 'inference' functions. It was at the level of utilizing intelligence generated in the cloud in the field. However, ROHM's On-Device Learning performs both 'learning' and 'inference' on the device itself. This is the biggest differentiator. Thanks to this, even after installation in the field, AI models can be continuously updated (retrained, additionally trained) in response to environmental changes or machine obsolescence. In other words, it becomes possible for the AI to adapt to and evolve in the field.

• AI Algorithm Solist-AI™: The algorithm underpinning this innovation is Solist-AI™. This technology is based on the '3-Layer Feedforward Neural Network (FFNN),' which features a relatively simple structure among neural network models that mimic the human brain. It applies a 'modified ELM (Extreme Learning Machine)' algorithm optimized for on-device learning. This algorithm employs a unique approach that updates only the weights (𝜷) connecting the intermediate layer and the output layer, while not updating the weights (𝜶) connecting the input and intermediate layers of the neural network. By training only the other side while keeping one connection fixed, it dramatically reduces the computational load, serving as a key factor in enabling fast and efficient learning even in MCU environments with limited resources.

• Proprietary AI Accelerator AxlCORE-ODL : AxlCORE-ODL is a hardware AI accelerator that embodies ROHM's technological ingenuity. This accelerator is designed to function as a peripheral device that operates independently of the MCU's CPU. When sensor data is input, the CPU simply instructs AxlCORE-ODL to begin AI processing. Subsequently, AxlCORE-ODL handles all complex computations for training and inference; it even directly processes demanding data preprocessing tasks, such as the Fast Fourier Transform (FFT), in hardware, thereby freeing up the CPU. Thanks to this structure, AI processing places almost no load on the CPU, enabling processing speeds theoretically about 1,000 times faster and ultra-low power operation compared to software processing.

Through this combination of an efficient algorithm called Solist-AI™ and powerful hardware called AxlCORE-ODL, ROHM has implemented a completely standalone AI solution using only a single small MCU chip. Now, let's look at specific examples of how this technology is changing actual industrial sites.

The Role of AI MCUs Changing Industrial Sites

The true value of technology lies in how it solves problems and creates new value in real-world applications. ROHM’s on-device learning AI MCUs have already proven their potential across various industries, accelerating the future of smart factories.

• Predictive Maintenance: Detecting Failure Signs in Factory Motors. This process is a prime example of 'unsupervised learning.' Instead of learning in advance what constitutes a 'defect' signal, AI learns only the complex patterns of the 'normal' state and detects all deviations from them. Thanks to this, it is possible to detect even unexpected types of failures. Imagine a motor, the core of factory automation. ROHM's AI MCU receives 3D vibration data in real-time from an accelerometer attached to the motor and independently learns the patterns of the 'normal state.' Subsequently, as the motor continues to operate, if subtle abnormal signs such as bearing damage or load imbalance occur, the AI detects the difference from the learned 'normal' pattern and outputs an abnormality index indicating 'deviation from normal.' Based on this signal, managers can replace parts or repair the motor before it completely fails. This is a case that perfectly implements 'Condition-Based Maintenance (CBM),' which goes a step further than reactive maintenance—repairing after a breakdown occurs—by accurately predicting maintenance timing based on the condition of the equipment.
• Degradation Prediction: Predicting the Lifespan of Photoelectric Sensors Photoelectric sensors, which detect the presence or absence of parts on a production line, experience performance degradation in their light-emitting and light-receiving sections over time. ROHM's AI MCU learns light intensity data from the light-receiving section along with the sensor's cumulative measurement count. Through this, it identifies performance degradation patterns and generates a 'degradation index' that predicts the sensor's remaining lifespan. This enables managers to efficiently manage maintenance schedules by individually optimizing the replacement cycle of each sensor.
• Ultrasonic forensics that detects even microcracks. Going beyond standard vibration detection, ROHM’s MCU can process data from Acoustic Emission (AE) sensors that detect weak high-frequency elastic waves generated when microcracks occur in materials. This captures defects at an extremely early stage that cannot be detected by vibration or temperature, representing a cutting-edge application that elevates predictive maintenance to the realm of 'microscopic forensics.'
• Expansion into various industrial fields. ROHM's technology is not limited to specific fields. It can be applied wherever sensor data is generated, such as detecting anomalies in factory automation (FA) sensors, determining the optimal adjustment time for industrial robots, predicting failures in home appliances, and monitoring the load status of power tools. As immediate decision-making becomes possible at the site where data is generated, new opportunities for value creation that were previously difficult to imagine are opening up.

These successful application examples clearly demonstrate that ROHM's AI MCUs can go beyond simple components and become game changers that fundamentally alter the way industrial sites operate.

The Future Opened by 'Self-Thinking Chips'

When considering the technologies and examples examined so far, the industrial implications of ROHM's on-device learning technology extend beyond simply 'smaller and faster AI.' This signifies a fundamental paradigm shift regarding the role and position of AI and will become a key driving force in establishing ROHM's unrivaled position in the market.

This signals a fundamental shift in industrial IoT business models. ROHM is not simply selling smarter chips; rather, it is providing key tools that enable equipment manufacturers to transition beyond machine sales to 'guaranteed uptime' or 'Performance-as-a-Service' models. The 'completely standalone' characteristic of this chip is the lynchpin that makes all of this possible by separating service value from cloud infrastructure costs.

The differentiating factors of ROHM technology compared to competitors' solutions can be summarized in three keywords.

1. 'No need for cloud pre-training': While most edge AI cannot completely escape cloud dependency, ROHM's solution enables zero-base training in the field.
2. 'On-site relearning possible': Rather than being a one-time installation, the AI adapts and evolves autonomously to the changing environment. This enables the maintenance of optimal performance throughout the equipment's entire lifecycle.
3. 'Completely Standalone': The fact that it requires no network connection at all provides an overwhelming competitive advantage in terms of security, cost, and ease of installation, and will be a powerful weapon, particularly for targeting the retrofitting market for existing facilities.

These three characteristics combined provide ROHM with a strong competitive advantage (moat) in the 'retrofitting' market, which adds intelligence to existing industrial equipment that is particularly large and historically difficult to penetrate.

In conclusion, ROHM did not simply create an AI semiconductor chip, but paved the way to implant 'self-learning intelligence' into all devices. AI has now begun to permeate every circuit around us, extending beyond massive data centers.

The era when AI permeates circuits

ROHM's on-device learning MCU has set an important milestone in the direction AI technology should take. AI is no longer a massive brain located far away, but is evolving into an entity that permeates the nervous systems of all objects around us, reacts instantly, and learns on its own.

When the era of AI permeating circuits fully begins, the future of manufacturing, robotics, smart homes, and all the electronic devices we use will fundamentally change. Machines will no longer be entities that move passively according to fixed programs, but will become intelligent partners that independently perceive their environments and conditions to make optimal decisions. A world where failures are predicted in advance, performance degradation is self-diagnosed, and operations are optimized to match user patterns is just around the corner.

A future where technology becomes smaller, smarter, and closer to us. ROHM's small semiconductor chips are leading a quiet but powerful revolution at the center of this massive change.