Embedded AI: from connected IoT to smart IoT
For years, many IoT devices have been limited to capturing data, carrying out basic processing, and sending or managing it according to the requirements of each product. This approach remains valid and effective. In fact, numerous devices now co-exist with us in an efficient and controlled manner.
However, technology is advancing. And, just as people reinvent themselves and adapt, IoT devices must also evolve.
Cloud-based artificial intelligence platforms and services are well known. They range from APIs specialising in machine learning to generative solutions such as ChatGPT, Gemini, Claude and Copilot. The question seems obvious: Why not use them to process data and assist with the device’s internal management?
The answer lies in the very nature of the IoT. Such equipment is usually subject to very specific requirements: low energy consumption, reasonable costs and efficient communication capabilities. Constant reliance on the cloud can introduce latency, increase power consumption, raise communication costs and pose privacy risks.
This is where embedded AI comes into play: the integration of machine learning models directly into the device’s microcontroller or low-power processor.
When used in microcontrollers with particularly limited resources, the term ‘TinyML’ is often used
We are not referring to cloud-based AI platforms, conversational agents or general-purpose generative models. We’re talking about small models, optimised and trained to perform a specific task.
Instead of sending all the raw data to external services, the device itself can analyse it locally and make immediate decisions. This allows you to take advantage of intelligent processing without having to rely on the cloud all the time.
What do we mean by embedded AI?
Embedded AI is a specific form of artificial intelligence, designed to respond to the context and the needs of the device. Its aim is not to respond to any request, hold a conversation or generate content, but to carry out a specific task based on the data collected by the system itself.
It can detect an anomaly, classify a signal, recognise a movement pattern or identify a relevant event.
Thanks to this technology, many IoT devices are no longer limited to simply capturing and sending data. They begin to interpret them right where they are generated. An industrial sensor, for example, can identify abnormal vibration patterns before a fault occurs. A wearable device can detect a fall or a sudden change in the user’s activity. A smart camera can filter out relevant events without streaming video continuously. And an environmental sensor can distinguish between normal fluctuations and situations that require an alert.
In all these cases, artificial intelligence enables decisions to be made locally, reduces reliance on the cloud, and allows for a faster and more efficient response. This approach also helps to reduce latency, energy consumption and network traffic. Furthermore, it enhances privacy, as not all data needs to leave the device.
From the model to the embedded application
The process begins with the collection of data from the context and environment in which the application will operate. Information can be obtained from accelerometers, gyroscopes, microphones, environmental sensors, electrical signals or any other relevant source.
Next comes the preparation phase, which may include cleaning, labelling, segmentation and feature extraction. In some cases, the aim will be to distinguish between normal and abnormal behaviour. In other cases, classifying signals, recognising patterns or estimating a value. In any scenario, the model must be as compact and efficient as possible.
The training then takes place. For this type of application, there are tools available that simplify the workflow:
- Edge Impulse: integrates data management, training, optimisation and the deployment of Edge AI solutions.
- NXP eIQ Time Series Studio: enables the development of time-series models and the generation of libraries for NXP devices.
- LiteRT / TensorFlow Lite: provides tools for converting and running compatible models on microcontrollers.
Once trained and validated, the model is exported in a format suitable for the embedded environment. Depending on the tool used, the result may be a model compatible with LiteRT or TensorFlow Lite, a library generated using eIQ Time Series Studio, or an SDK exported from Edge Impulse.
Use cases
The possibilities are endless. In industry, embedded AI can be used for predictive maintenance, anomaly detection or classifying the different operating states of a machine.
In the field of health and wellbeing, it can be incorporated into wearables capable of detecting falls, activity patterns or physiological events. In logistics, it enables the monitoring of transport conditions. In smart cities, it can analyse environmental or acoustic signals without the need to continuously send raw data to the cloud.
All these cases share a common theme: the device is no longer merely a tool for capturing information, but begins to interpret its context. It can achieve a level of accuracy that is difficult to attain using a traditional programming approach, thanks to model training and, where necessary, external retraining with new data.
Ultimately, embedded AI enables IoT devices to evolve from simply collecting data to systems capable of interpreting their environment and acting locally. A leap that paves the way for faster, more efficient solutions that are ready for the technological future.