ST has just formally announced the Edge AI Suite, a development tool set for its line of STM32 microcontrollers based on the Arm Cortex-M processor core, Arm-based Stellar automotive microcontrollers, and associated smart sensors. The Edge AI Studio is free to use when generating code for ST’s microcontrollers. ST’s NanoEdge AI Studio, one of the tools within the Edge AI Suite, can be licensed for use with Cortex-M microcontrollers offered by other semiconductor vendors. The Edge AI Suite and its associated libraries are not designed to implement large language models (LLMs) or for replicating cloud-based chatbots like ChatGPT inside of an air fryer, a toaster oven, or a washing machine. Instead, these tools are designed for developing edge and embedded applications where machine learning (ML) tasks can significantly improve the performance or feature set of an end product.

If you’re like me, you might have some difficulty in envisioning just what such applications might be. In a December presentation where the company rolled out its new tool suite, ST provided examples that I found very helpful in visualizing these applications. The first example was a microcontroller-based washing machine. The microcontroller is already incorporated into the washing machine’s design for motor control, with algorithms based on the STM32 motor-control SDK. Using ML techniques with no additional hardware, the microcontroller can determine the mass of the clothes placed in the washing machine by spinning the washing drum and using an ML algorithm to convert the motor’s feedback voltage and current into a clothing mass estimate.

Using this technique, the microcontroller can determine the weight of the clothes within 100 grams, which is three times better than existing techniques. The mass of the clothes establishes the precise amounts of energy, water, and detergent needed to clean the clothes thoroughly, again using ML model calculations. The result is clean clothes with lower energy costs, less water consumption, and less detergent pollution in the resulting wastewater. According to ST, the addition of ML modeling reduces the washing machine’s energy and water consumption by as much as 40%. Motor current consumption during the wash cycle can also be used for predictive maintenance using another ML model.

Yet another ML model takes streamed data from an ST 6-axis motion sensor in the washing machine to detect the position of the rotating washing drum. That information is used for collision avoidance, which prevents the spinning drum from hitting the stationary parts of the washing machine. All the ML models and software enhancements fit into the existing on-chip memory in the washing machine’s STM32G0 microcontroller, so there’s no incremental hardware cost for adding most of the ML models to this application. The one incremental hardware cost is the addition of the 6-axis sensor.

In addition, these ML tasks need no cloud connectivity. The washing machine can operate independently from the Internet, although connected appliances are a growing category as appliance manufacturers seek to bind customers ever more closely to their appliance brands to sell them accessories, consumables, maintenance, and other services. My relatively new Samsung Smart TV needs to call back to the Samsung mothership when it boots up, to see what features I’m entitled to use. Unfortunately, if it cannot contact Samsung, I get the equivalent of the Blue Screen of Death, which is a reminder that any connected appliance needs a fallback operating mode when it loses connection to the Internet. ST mentioned that a “major home appliance maker” will be introducing a product based on this ML technology some time in 2024.

The second application that ST mentioned was the use of ML techniques to determine rotor temperature in an electric vehicle’s traction motor while under load. The rotor temperature is an important measurement and is needed to operate the motor within safe limits and is also useful for predictive maintenance. In the lab, it’s easy to instrument a motor by opening an inspection cover to measure rotor temperature continuously under different voltage, current, and load conditions. It’s not so easy to measure rotor temperature in the field where the motor must be sealed against the elements. In this application, ST developed an ML model in the lab that infers rotor temperature from the motor’s operating conditions and from temperature readings from its outer enclosure.

That ML model running on a Stellar automotive microcontroller then serves as a virtual motor rotor temperature sensor in the running vehicle, which provides vital information to the motor-control algorithms on the road. The same microcontroller uses another ML algorithm for detecting anomalous motor behavior through vibration analysis. ST points out that the same sort of virtual temperature sensor can be created for the vehicle’s batteries using similar ML modeling techniques. Of course, the technique of using ML models to create virtual sensors is not limited to automotive applications. The technique should be broadly applicable to a range of real-world control applications.

The third application that ST mentioned was the implementation of an ML algorithm directly on an ST 6-axis motion sensor, which is used in HP laptop computers to monitor user activity. The ML application can sense when the laptop computer is set on a table or a lap with a user actively typing or clicking away, when the laptop is placed in a bag, and when it’s taken from the bag. The smart sensor then uses the results from the ML model to set the laptop’s operational mode to conserve battery power and stretch the interval needed between battery charges whenever possible. ST noted that the sensor drew only 34 microamps while performing this function.

The newly announced Edge AI Suite collects several existing ST tools and adds ML-specific components. The existing ST development tools included under the Edge AI Suite umbrella include the NanoEdge AI Studio for creating ML model libraries, STM32Cube.AI for optimizing and deploying trained Neural Network models from the most popular AI frameworks on STM32 microcontrollers, an OpenSTLinux expansion package that contains Linux AI frameworks and targets STM32MP1 series microprocessors, MEMS-Studio for developing no-code ML algorithms for ST’s MEMS sensor portfolio, and StellarStudio for developing and deploying embedded applications for ST’s Stellar automotive microcontrollers. The Edge AI Suite is compatible with existing compilers, assemblers, and debuggers for ST’s microprocessors and microcontrollers. On the front end, ST’s Edge AI Suite includes a library of ML use cases and an ML model zoo. It also includes access to ST’s Edge AI Developer Cloud, which ST announced early in 2023. The Edge AI Developer Cloud provides access to the models, libraries, and tools listed above as well as sample code.

ST’s NanoEdge AI Studio, a PC-based AI development studio that creates tinyML libraries with the push of a button, is available for use without cost when generating code for ST microprocessors and microcontrollers. The company is so sure of this tool that it also plans to offer paid license access to the suite for generating code to be used on other semiconductor vendors’ microprocessors and microcontrollers that incorporate Arm Cortex-M processor cores. Of course, that paid access will also give ST the opportunity of trying to sell its microcontrollers into these new ML-using processor sockets, so there’s a real benefit to ST for this largesse. Also, I note that some FPGAs also incorporate hardened Arm Cortex-M processor cores – including FPGAs from Intel and Microchip – and it’s possible to drop a soft Cortex-M processor onto almost any FPGA’s programmable logic fabric, which offers tantalizing possibilities for these ST ML development tools, if ST will support that sort of use.

Finally, ST referred to a new STM32 microcontroller called the STM32N6, which combines an Arm Cortex-M processor core with a “Neural-Art Accelerator,” a built-in neural processing unit (NPU) that can accelerate ML model execution. ST has been talking about this new microcontroller for more than a year but has yet to disclose many technical details. An ST blog suggests that the Neural-Art Accelerator can boost ML model execution speeds by one to two orders of magnitude while offloading most of the ML execution burden from the STM32N6’s Arm Cortex processor. ST plans to officially launch the STM32N6 microcontroller in 2024 and has likely sampled the device to about a dozen lead customers already.