Key Takeaways
The IEEE Spectrum webinar explores how AI-based virtual sensors can estimate signals that are difficult or costly to measure, such as battery state of charge (SOC) in Battery Management Systems. By integrating AI models into system-level design, developers can optimize models for memory footprint and execution speed while ensuring performance and deployment constraints are met.
Background: The Rise of Virtual Sensing in Industrial Systems
For decades, physical sensors have been the backbone of monitoring and control in industrial systems. From temperature gauges in chemical plants to pressure sensors in automotive engines, direct measurements powered real-time decision-making. But physical sensors come with limitations — they wear out, require calibration, add weight and cost, and sometimes can’t be placed where needed. In electric vehicles, for example, measuring battery state of charge directly is impossible without invasive techniques that compromise safety and lifespan.
Virtual sensors emerged as a response. The idea isn’t new — model-based estimation has been used since the 1960s in aerospace and defense, most famously with Kalman filters estimating spacecraft trajectories. But those models relied on precise physics-based equations, which are hard to derive for complex, nonlinear systems like lithium-ion batteries or powertrain dynamics.
That’s where AI steps in. Instead of hand-crafted equations, AI models learn patterns from data. By training on historical sensor readings, temperature profiles, current draw, and known SOC values, neural networks can predict SOC with high accuracy — without adding a single physical sensor. This shift from physics-first to data-first modeling began gaining traction in the late 2010s, especially as embedded processors got powerful enough to run lightweight neural networks.
Industries like automotive, energy storage, and industrial automation started experimenting. Tesla, for instance, has long used software-based estimation to manage battery health and range. In wind turbines, virtual sensors estimate blade stress using vibration and wind speed data, reducing maintenance costs. The IEEE has tracked this trend closely, publishing standards on model-based design workflows as early as 2015, setting the stage for tools that unify simulation, AI, and deployment.
Integrating AI Models into System-Level Design
Using a practical example, the session demonstrates how AI models can be integrated into system-level design and validated against performance, resource, and deployment constraints. This involves designing, verifying, compressing, and deploying AI-based virtual sensors to embedded processors within a single environment.
The workflow begins with data collection — real-world sensor logs from a target system, like a battery pack under various charge and discharge cycles. That data feeds into a neural network training pipeline, often using MATLAB or Python frameworks. But the real challenge isn’t training — it’s integration. Most AI workflows end at the model file. Here, the model must become part of a larger system: the battery management unit, which also handles cooling, charging logic, and safety checks.
That’s why the webinar emphasizes a unified environment. Instead of exporting a model and writing custom C code, developers import the trained network directly into Simulink. There, it becomes a block in a system diagram, connected to other components like voltage regulators, thermal models, and communication interfaces. Simulations run the entire system — not just the AI — under edge cases: extreme temperatures, sudden load spikes, sensor dropouts. This exposes issues early, like a neural network that drifts when battery impedance changes over time.
The benefit? A closed-loop validation process. If the virtual sensor predicts SOC within 2% of actual values across 100 simulated drive cycles, the model passes. If not, developers tweak the architecture, retrain, and retest — all without touching hardware. That’s a shift from build-test-break cycles to predict-validate-deploy workflows.
Designing and Verifying AI Models
The workflow shows how to design, verify, compress, and deploy AI-based virtual sensors to embedded processors within a single environment. By integrating AI models into Simulink, developers can simulate and validate the behavior of neural networks and assess their performance in real-world scenarios.
Design starts with architecture selection. For battery SOC, a long short-term memory (LSTM) network is often chosen because it handles time-series data well — remembering past current draw and temperature trends. But LSTMs are computationally heavy. The webinar walks through alternatives: gated recurrent units (GRUs), which are lighter, or even feedforward networks with sliding windows if memory is tight.
Verification is where many AI projects fail. A model might hit 95% accuracy on training data but behave unpredictably when deployed. The session introduces formal methods — not just testing, but proving bounds on output. For example, the SOC estimate should never exceed 100% or drop below 0%, even with noisy inputs. Using tools like Simulink Design Verifier, developers can generate test cases that stress the AI under worst-case inputs and confirm these safety constraints hold.
Another verification step is strongness testing. What happens if the temperature sensor fails and reports -40°C? Does the AI hallucinate a frozen battery or default to a safe estimate? By injecting faults into the simulation, developers can check if the system fails gracefully. This level of rigor is non-negotiable in automotive or medical applications, where AI errors can have real-world consequences.
Optimizing AI Models for Embedded Deployment
The webinar highlights the importance of optimizing AI models for embedded deployment, where memory footprint and execution speed are critical factors. By applying formal verification techniques, developers can ensure that AI models meet performance and deployment constraints while minimizing memory usage.
A trained neural network might use 50 MB of memory — fine for a server, but impossible on a microcontroller with 256 KB of RAM. Optimization bridges that gap. The session details techniques like quantization, where 32-bit floating-point weights are converted to 8-bit integers. This shrinks model size by 75% and speeds up inference, since integer math is faster on most embedded chips.
Pruning removes redundant neurons — those with near-zero weights — without hurting accuracy. In one example, a battery SOC model was cut from 12,000 parameters to 3,000 with only a 0.8% drop in accuracy. That’s a win when every kilobyte counts.
Another technique is code generation. Instead of relying on AI inference libraries like TensorFlow Lite, which add overhead, the webinar shows how to generate library-free C code. This code is lean — no unused functions, no dynamic memory allocation — and tailored to the target processor. A Cortex-M4, for instance, benefits from ARM’s CMSIS-NN optimizations, which the code generator can automatically include.
The result? A virtual sensor that runs in under 5 ms per inference on a $3 microcontroller, using less than 10 KB of memory. That’s deployment-ready for mass production.
Evaluating Design Tradeoffs
Developers will learn how to evaluate design tradeoffs across accuracy, performance, and deployment targets. This involves generating and profiling library-free C code for embedded deployment and assessing the impact of AI model complexity on system performance.
Every choice has a cost. A deeper network might boost accuracy by 1.2%, but increase inference time from 4 ms to 9 ms — pushing it past the control loop deadline. The webinar provides a framework for comparing options: plot accuracy vs. memory, latency vs. power consumption, strongness vs. development time.
One case study compares three SOC models: a full-precision LSTM, a quantized GRU, and a compressed feedforward network. The LSTM achieves 96.5% accuracy but needs 45 KB and 8 ms. The GRU hits 95.1% with 18 KB and 3.2 ms. The feedforward model scores 93.7% with 8 KB and 1.5 ms. For a high-end EV, the GRU makes sense. For a low-cost e-bike battery, the feedforward model wins.
Profiling tools help make these decisions. Developers run the generated C code on a virtual processor model, measuring cycle count, RAM usage, and cache misses. They can simulate how the AI affects real-time performance — does it block other tasks? Does it spike power draw?
This data-driven approach replaces guesswork. Teams can present clear tradeoffs to stakeholders: “We can save $0.15 per unit by using a smaller model, but SOC accuracy drops by 1.8% in cold weather.” That transparency speeds up approval and reduces last-minute rework.
What This Means For You
Developers and researchers in the field of AI and machine learning can benefit from this webinar by learning how to design and integrate AI models into system-level design. By optimizing AI models for embedded deployment, they can create more efficient and effective AI-based systems that meet performance and deployment constraints.
Consider a startup building smart HVAC systems. They want to estimate indoor air quality without adding $50 in gas sensors. Using the webinar’s approach, they train a virtual sensor on temperature, humidity, CO2 (from a single sensor), and occupancy patterns. They simulate the model in Simulink with real building data, verify it doesn’t false-alarm during cooking events, then deploy a 6 KB model to their existing controller. No hardware changes, no BOM increase — just smarter software.
Or imagine an automotive supplier developing a 12V battery monitor for stop-start systems. Physical voltage sensors degrade over time. A virtual sensor, trained on cranking current and ambient temperature, predicts battery health and flags failures before they strand drivers. The team uses the workflow to compress the model to fit on an 8-bit MCU, cuts development time by 40% with automated testing, and ships a product that outperforms competitors relying on older estimation methods.
For university researchers, the value is in reproducibility. Too many AI papers show lab results that never make it to hardware. This workflow forces early consideration of deployment — if a model can’t run on a $2 processor, it’s not useful for real-world applications. Students learn not just to train models, but to build them into systems, test them under stress, and generate production-ready code.
Competitive Landscape and Adoption Trends
While the webinar focuses on workflow, its implications ripple through the competitive landscape. Companies that master AI-based virtual sensors gain an edge: lower component costs, faster time to market, and higher system reliability. Legacy vendors relying on dense sensor arrays face pressure to adapt or lose share.
In the EV market, where every dollar and kilogram counts, virtual sensors are becoming a differentiator. Startups like Canoo and established players like BMW are reducing physical sensors in favor of software-defined estimation. That shift favors teams with strong model-based design skills — not just data scientists, but systems engineers who understand control theory and embedded software.
Tooling is also evolving. MathWorks, a key player in this space, has integrated AI into its core Simulink platform, making it easier for non-AI specialists to adopt. Competitors are responding: Siemens offers similar workflows in its Simcenter suite, while open-source tools like Scikit-learn and ONNX are being adapted for embedded use. But the gap remains — few platforms unify AI training, system simulation, and code generation as tightly.
Adoption is accelerating in regulated industries. Medical device companies use virtual sensors to estimate blood glucose from heart rate and activity, reducing the need for invasive monitors. In aerospace, AI models predict engine wear from vibration and exhaust data, enabling predictive maintenance. These applications demand the kind of verification the webinar emphasizes — not just performance, but provable safety.
Key Questions Remaining
Despite progress, open questions remain. How do virtual sensors handle concept drift — when real-world conditions change over time, making the model obsolete? Retraining is an option, but embedded systems rarely get updates. Some teams are exploring online adaptation, where the model tweaks itself using new data, but that introduces risks: what if it learns the wrong pattern?
Another issue is trust. Engineers are used to sensors with datasheets — known accuracy, failure modes, warranties. An AI model is a black box. Even with verification, explaining why it made a specific decision is hard. Regulatory bodies are starting to ask for audit trails: not just what the model predicted, but why.
Finally, who owns the virtual sensor? Is it part of the ECU software, the AI platform, or the system design? As these models become mission-critical, intellectual property and liability questions will grow. A faulty SOC estimate that leads to a battery fire could spark legal battles over whether the fault lies with the data, the algorithm, or the integration.
The webinar doesn’t answer all these questions — but it equips developers to build systems that are strong, efficient, and ready for the next wave of embedded AI.
Forward-Looking Question
As AI-based virtual sensors become increasingly prevalent in industrial and consumer applications, how will this impact the design and development of AI models in the future?
Sources: IEEE Spectrum, Technology Review
