The best sensor is the one you don't have to build — or buy, or calibrate, or replace when it breaks.
The Summary
- Wiley's Knowledge Hub webinar showcases a full workflow for designing AI-based virtual sensors that run on embedded processors using MATLAB and Simulink
- Virtual sensors use AI models to infer physical measurements (temperature, pressure, position) from other available data, eliminating hardware costs and failure points
- The workflow includes formal verification of neural network behavior and generates library-free C code for deployment, addressing the reliability gap that's kept AI out of industrial systems
The Signal
Industrial systems are sensor-heavy by necessity. Measure temperature here, pressure there, vibration everywhere. Each sensor adds cost, weight, wiring, and a new thing that can fail. This workflow from MathWorks flips that equation: train an AI model to infer what a physical sensor would measure using data you already have.
The concept isn't new. Virtual sensors have existed in control systems for years using physics-based models. What's different now is using neural networks instead, which can capture complex nonlinear relationships that traditional model-based approaches miss. The tradeoff has always been trust. You can mathematically prove a Kalman filter's behavior. A neural network is a black box.
"Apply formal verification techniques to assert neural network behavior."
That verification piece matters more than the AI itself. In automotive, aerospace, medical devices, you can't deploy a model that might hallucinate a temperature reading. The workflow lets engineers set bounds on network behavior, verify those bounds mathematically, then test the deployed code against the verification claims. It's the difference between "this works in testing" and "this works, and here's the proof."
The compression step addresses the other deployment barrier: embedded processors have tight memory budgets. A model trained on a workstation might be 50MB. The compressed version for a microcontroller might be 500KB. You're trading some accuracy for a model that actually fits where it needs to run.
Key capabilities:
- Library-free C code generation (no runtime dependencies to debug)
- Processor-in-the-loop testing to catch deployment issues before production
- System-level simulation that includes the virtual sensor, the control algorithm, and the physics
The workflow is end-to-end in the sense that matters to engineers: design, verify, compress, deploy, test. All in one environment where the model doesn't get lost in translation between tools. For industries where sensor failures mean downtime (manufacturing) or danger (automotive), virtualizing sensors isn't about cost savings. It's about eliminating failure modes.
The Implication
Virtual sensors won't replace all physical sensors, but they'll replace the expensive, hard-to-place, or failure-prone ones first. Watch for this in automotive (fewer sensors per vehicle, lower cost), industrial IoT (inferring equipment health from indirect measurements), and anywhere weight matters (aerospace, robotics). The companies that figure out verified AI sensor models will have a reliability advantage their competitors can't match with more hardware.
If you're building embedded systems, the question isn't whether to use virtual sensors. It's which ones to virtualize first and whether you can prove they work before something breaks.