Is Embedded Machine Learning (ML) Becoming the Norm for Industrial Edge Applications?

The Industrial Internet of Things (IIoT) generates vast amounts of data, yet only a small portion is utilized. Over the past 5 years, Artificial Intelligence (AI)-based cloud computing has unlocked the value of this underutilized data, offering more efficient data integration and analysis. Companies are increasingly recognizing the benefits of embedded Machine Learning (ML) for data filtering and pre-processing, leading to lower cloud storage costs, better data structuring from industrial edge equipment and sensors, and enhanced autonomy for edge operations.

Embedded Machine Learning Market

Sensor-based monitoring is rapidly growing in the embedded ML IIoT sector, with significant growth projected for Condition-Based Monitoring (CBM) applications. Embedded ML, particularly at the sensor level (TinyML), is expected to see rapid growth in the industrial edge due to investments from Microcontroller Unit (MCU) vendors and interest from Original Equipment Manufacturers (OEMs), with significant expansion anticipated from 2026 onward. Networking hardware, such as gateways and routers, will play a crucial role in edge ML architectures by aggregating sensor data for contextual analysis before cloud processing.

“Much of the innovation in embedded ML is at the ‘developer’ toolset level, as these companies look to improve the quality of embedded ML models across different axes, including model size, model accuracy, model stability, inference quality, and input signal range. Additionally, these vendors look to offer an easy-to-use interface, usually using AutoML+ tools, to simultaneously target embedded developers, data scientists, and device makers less versed in ML development.” – Tancred Taylor, Senior Analyst at ABI Research

Embedded ML Supporting New Industrial Edge Use Cases

To industrial companies, the idea of the “industrial edge” typically ends at the network layer/industrial computer. This assumption serves as the impetus for networking equipment being the primary focus for embedded ML, with use cases like defect detection/machine vision dominating the discussion. However, ABI Research has observed that embedded ML computing technologies are rapidly shifting to the machine and sensor edge.

Various vendors are optimizing large models built in standard ML frameworks for smaller hardware. Alternatively, vendors allow industrial edge users to train and build models from scratch, fine-tune the models for low-compute hardware, and assist the device deployment process. There will be another wave of embedded ML innovation, supporting use cases, and time series data analysis use cases, particularly in Condition-Based Monitoring (CBM) and predictive maintenance. In the future, our analysts anticipate embedded ML computing to be a staple in the wider industrial edge, including networking, edge computers, machines, and sensors.

You May Also Like:

Nokia Launches MX Grid, Bringing Processing and ML Closer to the Far Edge

TinyML Is A Perfect Match for Environmental Data

Embedded ML Development Has Become Easier, but Challenges Linger

Building and optimizing embedded ML models for different layers of the industrial edge has become relatively straightforward. Toolsets from cloud vendors like Amazon Web Services (AWS) and Microsoft cater to high-compute edge requirements, while developers such as Edge Impulse, SensiML, Stream Analyze, MicroAI, TDK (Qeexo), and STMicroelectronics (Cartesiam) provide mature tools for creating embedded ML software suited for more resource-constrained industrial edge applications.

Despite the ease of building these ML models, deployment and operationalization still present challenges:

• Ensuring data quality, quantity, and type from industrial edge machines.
• A lack of partnerships between horizontal toolset providers and domain-specific edge application platforms hinders solutions' operationalization.
• Industrial software vendors often focus on cloud solutions and are less familiar with embedded ML tools.
• The edge orchestration, observability, and operationalization tools of embedded ML applications below the industrial computer layer are relatively immature.

New vendors like MicroAI, Stream Analyze, and Barbara are emerging to address these gaps and enhance resource-constrained embedded ML deployment in industrial markets. The market for embedded ML consists of two main categories: developer-focused and deployment-focused. The next two sections of this article evaluate the emergent trends regarding embedded ML toolsets for these two categories.

Figure 1: Embedded Machine Learning Ecosystem for IIoT Monitoring

(Source: ABI Research)

Embedded ML Developer Tools

Below is a list of the latest trends in embedded ML developer-focused toolsets for the industrial edge:

• Time Series Data: Embedded ML for video and text processing is relatively easy to build and deploy, while time-series data (vibration, acoustic, temperature, etc.) is more challenging due to proprietary data and diverse sensor signals. Optimized solutions are emerging, leading to new sensor analysis use cases at the far industrial edge and on networking equipment.
• Smaller Models: There is a trend toward smaller models deployable on both high-performance hardware and less powerful devices like gateways and smart sensors. To support this facilitation, companies like Edge Impulse and SensiML facilitate deploying accurate models on resource-constrained hardware, with early developments like TinyML on Intelligent Sensor Processing Units (ISPUs) showing promise.
• Foundation Models: Embedded ML is consultative and costly for industrial companies, but foundation models offer an easier start, using field data for continuous learning. These models will evolve to focus on specific verticals or domains, improving stability and accuracy over time.
• Using AI for Data Pre-Processing and Labeling: AutoML tools, or "AI builds AI," automate model building, training, and optimization. The next step involves automating industrial data collection, cleaning, and labeling to ease IIoT adoption.
• Closer Relationship between Hardware and Software: Computing platform vendors are heavily investing in embedded ML toolkits, covering the entire development value chain (data ingestion, model zoos/libraries, etc.) and often offering these tools for free or at nominal prices. In turn, they are commoditizing ML building and training skills.
• Building Developer Competency: Developer tools are targeting OEMs and System Integrators (SIs), but expertise gaps exist. Embedded ML vendors are addressing this by supporting open-source communities, offering AutoML+ tools to lower entry barriers, and providing training to help build in-house expertise.

Embedded ML Deployment Tools

Bringing embedded ML to market for the IIoT necessitates two critical features often missing in developer toolsets: workflows/applications and orchestration capabilities.

• Scaling to Production: Moving from Proofs of Concept (PoCs) to production requires robust and securely embedded ML models, continuous learning orchestration, and access to real-time data from industrial Operational Technology (OT) field assets. Specific foundation models, domain-specific partnerships, and advanced Machine Learning Operations (MLOps) tools will help address these challenges.
• Vertical-Focused Scale Partners: Although it's now "easy" to build and optimize models for industrial edge devices, toolset vendors often lack domain-specific knowledge for IIoT markets. Partnerships with application platforms (e.g., Litmus Automation and PTC) and SIs will facilitate real-world deployment of AI/ML models.
• Orchestration and MLOps Are Key, but Primarily Focused on “Thick Edge”: Edge orchestration has mostly focused on cloud-network relationships, but embedded ML tools for thin-edge or sensor-edge deployments are less common. Companies like MicroAI and Barbara are developing platforms to operationalize embedded ML at the thin and sensor edge.
• Operationalization of Developer Toolsets: Effective embedded ML deployment requires industrial connectors, model observability, orchestration across edge and cloud, model versioning, decentralized learning, and standard integration into customer systems.

Key Companies

• AWS
• Edge Impulse
• Qeexo
• SensiML
• STMicroelectronics
• TDK

Learn More

Explore the latest trends, challenges, opportunities, and vendors in the world of embedded machine learning by downloading our comprehensive report: Embedded ML for Industrial IoT Monitoring and Maintenance. More research like this can be accessed by signing up for our Industrial & Manufacturing Technologies and IoT Markets research services.