Transforming software-defined vehicles with neural-style embedded design

By Yoann Foucher, Marketing Director, STMicroelectronics

As automotive markets adopt flexible, software-defined vehicles, AI lets engineers deliver high-value features that are too complex for traditional rules-based systems and too time consuming to develop using conventional tools. By adopting neural-style embedded design, teams can compress time to market for new products that deliver improved end-user experiences, while also easing compliance with safety legislation.

Two specific use cases demonstrate how AI running on an MCU can handle variables that are difficult to control properly across a wide range of operating conditions using traditional code. The first addresses the mechanical and electrical challenges associated with controlling a powered sliding-door for a commercial van or a passenger minivan. The second quickly and accurately identifies the signatures of electrical conditions that can lead to arcing in 48V vehicle wiring and modules, letting the system react to ensure safety.

Solving Real-World Physics with Edge AI
In the sliding door example, gravity becomes an unpredictable enemy when a heavy door is operated on steep, sloped terrain. If the vehicle is parked nose-up on a fifteen percent slope, the motor must fight an immense load to close the door. Conversely, if parked nose-down, gravity can cause the door to slam dangerously, risking mechanical wear, structural damage, or injury to passengers.

Developing an application to control the motor drive using traditional techniques requires engineers with specialized knowledge of physics and mechanics to create code that can handle all possible operating conditions. This includes evaluating every variable the system will encounter, such as if the vehicle is parked on any kind of gradient, and ensuring proper operation of safety features like anti-pinch. Traditional approaches require months of development, including manually writing and testing proportional-integral-derivative motor-control loops and building static look-up tables. This process is relatively slow, rigid, and difficult to scale across different vehicle models.

With neural processing, the development flow can change from conventional manual logic bring-up to data-driven training. Instead of engineers attempting to predict every possible scenario the door might encounter and writing a suitable mathematical formula, development instead focuses on data acquisition, recording motor behavior across various slopes, temperatures, and loads. An automated machine-learning pipeline then analyzes the data to select and optimize the best neural network architecture for the target hardware.

When the sliding door controller is designed for neural processing in this way, the application can predict the door’s kinetic energy before it reaches the latch. Adjusting the motor torque in micro-intervals can electronically brake the door as it slides downhill or deliver a boost to overcome a steep uphill gradient. This ensures a soft, consistent close each time, regardless of variables like the terrain or the weight of the vehicle’s contents. This transition from reactive coding to predictive AI-driven control transforms a simple mechanical component into an intelligent, adaptive feature that enhances both safety and the premium feel of the vehicle.

Predicting and Preventing Dangerous Conditions
The second challenge covers an important safety issue arising with the industry-wide adoption of 48V electrical architectures, which aims to provide more power for features like active suspension and steer-by-wire. With this move comes a dangerous physical reality in the form of stable electric arcs. Unlike 12V systems, where an arc usually extinguishes itself almost instantly, a 48V arc can sustain itself, generating temperatures high enough to melt insulation and initiate fires. Detecting these arcs using traditional threshold triggers with explicitly defined rules and filters is notoriously difficult. Standard electrical noise from components like wipers or cooling fans can sometimes closely resemble the signature of an arc, leading to frequent false positives or, worse, missed detections.

This is where AI excels through pattern recognition. An MCU-based AI model can see patterns in the high-frequency signature of the current and voltage on the 48V bus that no rule-based system can enumerate. The model can be trained to distinguish between a noisy motor and the specific, chaotic frequency spectrum of a developing carbonized path or a loose connector with unprecedented speed and precision.

The Need for Speed and the Case for Local Inference
A common misconception in the era of the Software-Defined Vehicle is that all intelligence should be centralized within the big brain of the car, typically a high-power Microprocessing Unit (MPU) or System-on-Chip. While this makes sense for infotainment functions like voice recognition and long-range path planning, it is less well suited to tasks that involve real-time physics and safety-critical control loops. The primary reason for this is communication latency. In a centralized automotive network, data must travel from a sensor, through a wiring harness, into a gateway, across a high-speed Ethernet or CAN-FD bus, and finally into the MPU’s memory for processing. The round-trip latency can exceed ten to fifty milliseconds.

In motor control or electrical safety, fifty milliseconds is an eternity. A sliding door traveling at half a meter per second moves twenty-five millimeters in that timeframe. A 48V arc can cause permanent structural damage in less than five milliseconds.

When AI inference runs locally on a microcontroller (MCU) located a few centimeters from the sensor pins, safety-critical actions can happen in real-time without network congestion or software hangs. In some automotive products, a local intelligence is provided by introducing a specialized Neural Processing Unit (NPU) called the Neural-ART Accelerator. This dedicated hardware block is architected for efficient data flow and parallel processing instead of the traditional sequential fetch-decode-execute cycle. By offloading complex models such as pattern analysis or multi-phase motor control to this neural accelerator, the main MCU cores can concentrate on standard control tasks. In this way, AI-accelerated MCUs can deliver as much as thirty times greater processing efficiency.

The arc-suppression system thus provides predictive safety, detecting the onset of an arc and tripping the electronic fuse before the heat reaches a combustion point. Moreover, the response time is guaranteed and repeatable. By implementing this intelligence at the edge, the vehicle gains a level of electrical surveillance that was previously impossible, making the 48V transition safer and more robust for mass-market adoption.

Local inference leveraging an integrated NPU like the Neural-ART Accelerator also supports the concept of virtual sensing, where AI models estimate internal parameters like battery health or motor temperature that are impossible to measure directly with physical sensors. This not only reduces the total cost of materials by eliminating redundant sensors but also reduces vehicle weight and wiring complexity, which are vital goals for modern electric vehicle design.

The MCU memory architecture also plays a pivotal role. In a Software-Defined Vehicle, memory is at a premium because the car must be able to store multiple software images and complex AI models while simultaneously handling over-the-air updates. AI models are known for their voracious demand for data, both for training and when running inference. Some architectures, offer twice the density of traditional flash memory thereby providing generous storage for typical edge AI models. This also ensures that, as AI models evolve and improve over the vehicle’s lifespan, there is sufficient headroom to deploy updates without hardware limitations. A car delivered today can thus become smarter and safer five years down the line through software alone.

Optimized Edge AI Ecosystem
The neural development flow benefits from an ecosystem that includes pre-trained, automotive-grade AI models. By providing a starting point for common tasks, these make it possible for teams to deploy sophisticated AI without needing a massive department of data scientists.

Moreover, this AI-driven approach is a powerful enabler for the industry’s move away from distributed domain-based architectures, where every function has its own dedicated box, toward zonal architectures. In a zonal setup, MCUs act as local hubs for sensors and actuators in a specific area of the car, eliminating many kilograms of copper wiring and reducing the number of separate ECUs.

Conclusion
The integration of AI onto automotive microcontrollers lets engineers solve complex physics problems like sloped-terrain door control and 48V arc detection at the edge. Leveraging this transformational change, the automotive industry can advance the concept of the software-defined vehicle to create vehicles that are truly intelligent. The microsecond-level response time afforded by local inference ensures that safety is never compromised by network latency, while the shift to a data-driven development flow allows OEMs to innovate at a pace previously reserved for the tech industry.