By George Dickey
Technology innovations often come in waves that change human life within a few decades: electricity, computers, the internet. The most recent wave is artificial intelligence (AI). Since its inception, AI has been predominantly limited to large computational platforms. However, the convergence of advanced processor technology and high-efficiency AI networks has led to groundbreaking innovations that allow AI to run within embedded systems. These systems are often equipped with specialized AI-specific processors and machine learning-enabled sensors that enable never-before-seen capabilities at the ‘edge.’
These capabilities allow new levels of predictive maintenance. Embedded AI acceleration can identify problems before they happen, without any human involvement. This article looks at several new processor technologies that enable AI algorithms at the edge.
Embedded AI systems
AI-enabled microcontrollers and MEMS sensors are at the forefront of the predictive maintenance AI revolution. These devices are characterized by their compact size, low power consumption, and ability to accelerate specific math functions related to AI. Traditional embedded processors, in combination with AI cores and/or sensor modules, enable devices to analyze and respond to real-world, time-series-based data in real-time. There are several ways embedded AI is being implemented in time-series-data applications. But first…
What is AI for time series data?
Time series data refers to a sequence of data points collected, recorded, or measured over evenly spaced intervals of time. Time series data points allow analysts to understand how data evolves or changes over time.
Time series data analysis involves understanding patterns, trends, anomalies, and behavior within data. AI can be used to make observations or predictions about future values, extracting insights from the data to inform decision-making. This type of analysis can be done using AI networks, which requires the understanding and selection of the processing hardware.
In applications such as predictive maintenance, environmental anomaly detection, IoT devices, multi-axis motion, and more, time series data can be used to understand patterns, trends, and behaviors within the data. Using AI algorithms such as convolutional neural networks, recurrent neural networks, long and short-term memory networks, and gated recurrent units, time series data can be used to detect desired or anomalous outcomes. While the execution of these machine learning algorithms can be done on generic hardware the use of processors and/or sensors with AI cores decreases the latency and increases the efficiency.
Several common processor core technologies, including Cortex-M cores, NPUs, GPUs, and embedded AI sensor assemblies, can be used for AI time series data analysis. The fusion of these new processor technologies with dedicated AI algorithms is driving innovation in embedded systems and edge computing. From applications in healthcare, automotive, manufacturing, agriculture, and beyond, embedded AI processors are paving the way for smarter, more autonomous devices that can analyze real-world data with unprecedented speed, accuracy, and efficiency.
Nanoedge AI Studio showing time series data from a motor control application.
Machine-learning capable microcontrollers
The Cortex-M series of microcontrollers (MCU), which range from M0 to M85, often serve as the backbone for embedded system processing in a variety of applications, whether AI is executed or not. However, because these cores are designed for low-power, real-time data processing, they are ideal for embedded AI hardware solutions.
For example, STMicroelectronics STM32L5 and NXP’s MCX-A which use 32-bit Arm Cortex-M33 are both suitable MCUs for use in embedded systems that utilize simple AI networks. While these traditional Cortex-M cores excel at handling sensor data and simple AI processing, for more complex machine learning tasks let’s look at microcontrollers that integrate additional cores to further enable machine learning.
Graphics Processing Units (GPUs)
While primarily intend for improving 2D (and sometimes 3D) graphics performance, GPUs are increasingly being utilized alongside Cortex-M MCUs for embedded AI applications. These parallel processing units can be utilized for deep learning algorithms such as convolutional neural networks (CNNs) for tasks such as image recognition and object detection. For example, the STM32U5 boasts a Cortex-M33, and a NeoChrome GPU, making it suitable for HMI applications or embedded AI solutions in industrial, smart city, smart home, and IoT applications.
Neural Processing Units (NPUs)
Neural Processing Units (NPUs) are highly specialized cores that are optimized for accelerating neural network computations, which enables the programming to functionally learn and reprogram itself. These cores, often implemented alongside Cortex-M processors, can execute more complex neural network algorithms than standard Cortex-M cores are capable of in isolation.
For example, NXP’s MCX-N combines an Arm Cortex-M33 and custom eIQ neural processing unit. Alif Semiconductor’s Ensemble family are industrial-application-ready microcontrollers that combine an Arm Cortex-M55 CPU with dedicated edge AI acceleration made possible by an ARM Ethos-U55 Neural Processing Unit. The family is available with a single Cortex-M55 or dual Cortex-M55, single or dual Ethos-U55, and optionally one or two Cortex-A32 MPU cores.
By offloading AI tasks to NPUs, embedded systems can achieve real-time neural network inference while conserving power, size, and resources.
Sensors with embedded AI cores
As discussed, embedded AI applications often utilize a standard MCU for the computational side of data processing. However, new sensor technologies have moved the AI processing external to the MCU and put embedded AI processing cores within the sensor itself, referred to as machine learning cores (MLC) and intelligent sensor processing units (ISPUs).
Sensors with an embedded machine learning core (MLC) can be trained to trigger actions when a specific event is detected, making it able to detect precise change scenarios. In doing so, computational load on the MCU can be reduced, which yields a low-power architecture and improves system efficiency. For example, LSM6DSV16BXTR is an IMU with a 3-axis accelerometer and 3-axis gyroscope that features an MLC to enable AI features.
Alternatively, sensors can feature intelligent sensor processing units (ISPU), which are integrated digital signal processors dedicated to high-processing capabilities to support machine learning and neural network processing within the ISPU. This core architecture allows for AI-powered processing of internal and external sensors, without needing an external MCU to handle the heavier computation. This is used for automatic calibration, sensor fusion, and anomaly detection across a variety of sensor inputs without needing an external MCU. Instead, smaller MCUs can be used for general-purpose microcontroller loads.
Conclusion
The application of AI to time-series data is an exciting area of development with the potential to add intelligence to industrial, healthcare, and consumer applications. There are many factors to consider in the development of an AI solution, the selection of the processor being just one of them.
For further information and guidance, you can book a consultation with Arrow AI experts, and follow our AI webinar series.
AI Consultation with Arrow experts
Related Products:
See related product
See related product
See related product
See related product
Additional AI Engineering Resources:
Alif Semiconductor
MCX General Purpose Microcontrollers
EMEA Vertical Autonomous Machines
EDGE AI Webinars
