MathWorks Demonstrated a Complete Development Cycle for AI Sensors in Embedded Systems

MathWorks announced a webinar on the complete workflow for creating and deploying AI models of virtual sensors that run directly on embedded processors—without cloud overhead and latency.

Why Virtual Sensors Are Needed

A virtual sensor is a neural network that computes a required measurement based on other sensors. For example, an air quality sensor calculated from pressure, humidity, and CO₂, or an equipment failure prediction system that analyzes vibration and temperature. In industry, such sensors have long been used physically—in thermal pipelines, hydraulic systems, and storage facilities. But in embedded systems, this is critical for different reasons: calculating in the cloud is expensive (the internet is unreliable, latency is unacceptable), and on a microcontroller, you must conserve every bit of memory and every milliampere of energy.

What MathWorks Demonstrates

The webinar reveals an end-to-end workflow that includes:

• Integration of AI models into Simulink for verification and simulation-based testing at the system level
• Formal neural network verification—a mathematical guarantee that the network operates within acceptable bounds
• Model compression (quantization, pruning) to save memory and accelerate execution
• C-code generation without dependencies on external frameworks
• PIL tests (processor-in-the-loop)—verification on the actual target processor
• Performance profiling and analysis of trade-offs between accuracy and speed

The entire cycle in a single environment—MATLAB and Simulink. No switching between tools, no loss of context.

Practical Challenges

The main problem: a neural network that works elegantly in a Jupyter notebook can fail on a microcontroller with 32 KB of memory and a 48 MHz processor. That's why MathWorks emphasizes formal verification—not just running tests, but proving that the network won't exceed the boundaries of safe behavior even in the worst case. The second bottleneck—how to compress the model so it fits in memory and runs fast enough, yet doesn't lose accuracy? There are techniques (quantization to int8, structural pruning, distillation), but without automation, it's manual labor.

Context: AI on Edge

This webinar is part of a major trend. Over the past 5 years, edge AI has moved from a niche to mainstream. Now not only large companies train models, but engineers who build washing machines, refrigerators, sensors, and pumps do as well. And they all face the same problem: how to run ML locally when the hardware is like that of the 2010s, but the requirements are those of 2025.

What This Means

Embedded AI is transitioning from hobbyist enthusiasm to production-grade engineering practices. If you have an IoT device, industrial equipment, or consumer electronics—in the coming years you'll need to embed AI models locally instead of sending everything to the cloud. MathWorks offers tooling that makes this an order of magnitude simpler—and most importantly, safer.