Architectural Considerations and a Modular Path Forward
Factory floor operators interact with complex machines through touchscreens, dashboards and displays that must adapt to safety requirements and production targets. At the same time, plants are juggling staffing shortages, aggressive productivity goals, and heightened security concerns.
AI‑enhanced human-machine interfaces (HMIs) can recognize individual operators, streamline workflows, reduce human error and enforce access policies. For instance, personalized user access through facial recognition can identify authorized operators via camera before permitting access to equipment or documentation. This reduces risk of unauthorized interactions and ensures only trained personnel operate critical machinery.
This AI‑driven capability is an incremental upgrade to existing HMI architectures and presents OEMs and systems integrators with a path to developing good-better-best industrial HMI product portfolios.
Understanding the best approach to designing such a portfolio starts with defining the basic architecture and components of a typical industrial HMI.
Architecture of a Standard Industrial HMI & AI Augmentation
A conventional industrial HMI is a complex system that typically runs an embedded operating system (like Linux or Windows), communicates with programmable logic controllers (PLCs) and sensors via serial ports or Ethernet, and presents operator screens created with an HMI/SCADA tool.
To support this functionality, a standard industrial HMI will include the following components, at minimum:
- A display (often between 7" and 15") with a touchscreen
- An embedded processor, memory, and storage
- A set of interfaces for fieldbus and Ethernet connectivity
- A rugged housing designed to operate in industrial environments with an ingress protection (IP) rating
Augmenting the conventional HMI architecture with AI capabilities requires increased resources. For example, facial recognition is based on convolutional neural networks that are trained to encode facial features. "Lightweight" models such as FaceNet or YOLOv3Face are available for these tasks and can be executed using OpenCV and LiteRT (TensorFlow Lite) to optimize inference on CPU and neural processing unit (NPU) hardware.
Performance-wise, delivering a smooth facial recognition user experience requires low latency and minimal memory overhead. While models can be lightweight (requiring about 1 MB RAM), they can demand at least 1 GB of system memory to accommodate multiple processes and secure storage of biometric templates.
Benchmarks from an example application on the SECO App Hub show that inference on a dedicated NPU delivers latency around 36 ms, while CPU‑only execution can exceed 440 ms. Clearly, different display computers are better suited to AI workloads than others, which means a discrete or integrated NPU or graphics processing unit (GPU) can help deliver real‑time performance without overloading the CPU.
Of course, additional components not listed in our conventional HMI architecture are also required to make this AI application a reality. They include:
- Camera module with sufficient resolution and lens acuity, as well as infrared (IR) capability for low‑light conditions. The camera is mounted in the bezel of the display and connected via a MIPI-CSI or USB port
- Edge AI acceleration using a processor with an integrated NPU, GPU, or vision processing unit (VPU) for offloading inference workloads
- ≥4 GB RAM and 16 GB eMMC for hosting the operating system, AI model, and application software
- Hardware root of trust, secure boot, and encryption to protect facial templates and authenticate operators
Design Approaches for AI‑Ready Industrial Display Computers
Automation equipment suppliers delivering industrial HMI product lines can now consider portfolios consisting of, for example, a conventional HMI, a personalized user access HMI, and more advanced versions with features like operator alert assistance.
There are a number of design approaches available to engineering teams looking to deliver such portfolios. Two of the most common are a monolithic design or a modular architecture approach:
- Monolithic designs tailor separate HMIs for each of the conventional display, personalized access, and advanced AI use cases. Each model in the product family leverages a dedicated processor and hardware configuration to optimize performance for the target application. However, the need to maintain multiple designs in parallel increases development cost and supply chain complexity.
- Modular architectures rely on computer-on-module (COM) standards to define the form factors and pinouts of off-the-shelf compute modules that plug into a carrier board. This carrier exposes all the interfaces required by a target application—Ethernet, USB, serial ports, camera, display, etc.—so developers can start from a single display and enclosure combination, then select from a range of COMs to scale from entry‑level CPU modules (basic HMIs) to high‑end AI processor modules (AI HMIs).
Image
By separating the embedded computing core circuitry from the I/O baseboard, SECO's Modular Vision HMI product family architecture allows processor upgrades without redesigning the entire HMI. Not only does this provide an upgrade path from low‑cost models to AI products, but computer on module standards like the Smart Mobility ARChitecture (SMARC) further reduce time‑to‑market and de‑risk processor selection because the same carrier can host Arm or x86 compute modules.
Modular Vision: A Flexible Approach to AI‑Ready HMIs
SECO uses a Modular Vision concept based on SMARC to align HMI display sizes, display computer processors, and the requirements of AI/non-AI applications. The table below compares three modular HMI models across key metrics such as display size, AI accuracy and latency, environmental tolerances, and cost.
| Modular Vision SKU | Display size (in) | Processor | AI performance (Personalized Access) | Environmental tolerance | Cost |
|---|
| Modular Vision 7 MX 93 | 7 | NXP i.MX93 (dual A55 with U‑65 NPU) | Moderate accuracy, ~150-200 ms due to Ethos U‑65; suitable for basic facial authentication; limited headcount per second. | IP65 front; 0 °C to 60 °C; 50 k hour backlight; 3 mm glass | Low -- entry‑level pricing |
| Modular Vision 10.1 MX 8M-Plus | 10.1 | NXP i.MX 8M Plus (quad A53 with 2.3 TOPS NPU) | High accuracy, real‑time latency (~36 ms) with NPU; supports multiple faces and voice commands. | IP65 front; 0 °C to 60 °C; 50 k hour backlight; 3 mm glass | Medium -- balanced cost and performance |
| Modular Vision 15.6 ASL | 15.6 | Intel Atom x7000RE (2/4/8 cores with UHD Graphics) | Very high accuracy using CPU with VNNI instructions and integrated graphics; can scale to multi‑face recognition; latency depends on core count (12-25 ms/20-40 ms/35-70 ms). | IP65 front; 0 °C to 60 °C; 50 k hour backlight; 3 mm glass | High -- premium performance and features |
Each of the above Modular Vision products uses the same baseboard and display options but swaps the compute module based on an application's performance requirements. OEMs can begin with the entry‑level 7" model for simple HMIs and later upgrade to the 10.1" or 15.6" variants or when adding advanced AI functionality.
Designing with Modular Vision
By adopting a SMARC‑based Modular Vision architecture, machine builders can utilize a scalable family of HMIs that share the same mechanical design, connectors and software environment. And as new processors or memory configurations are introduced, they can upgrade without re‑certifying the entire enclosure for environmental or safety compliance. When coupled with a modern Linux distribution such as Clea OS, the system provides remote device management, secure updates and integration with industrial IoT protocols. Application developers can deploy pre‑built AI apps—like the facial recognition app found on the SECO App Hub—using TensorFlow Lite, OpenCV, or PyTorch with confidence their algorithms will run across both Arm and x86 SMARC modules with only minor adjustments to performance parameters.
The models highlighted above demonstrate how screen size, compute capability, and cost can be aligned to support different AI‑enabled use cases and product lines. With rugged enclosures, industrial‑grade displays and standardized interfaces available off the shelf or custom from SECO, these systems can be manufactured and deployed at scale to bring AI to the edge of the production line.
Do you have an industrial application for an HMI? Contact SECO to discuss how Modular Vision, Clea OS, and AI algorithms can satisfy your latest product requirements.