Demystifying AI and ML with embedded devices

Machine learning (ML) enables electronic systems to learn from existing data and use their own knowledge to measure, predict, and make decisions. Since these types of applications normally require high computational power, they always run on PCs and cloud servers. Thanks to innovative techniques, algorithms and robust workflows, machine learning can now be carried out directly on embedded devices.

Benefits of ML on embedded devices for industries

ML on embedded devices has many advantages. It eliminates the need to transfer and store data on cloud servers, and reduces data loss and privacy leaks during data transfer. It also reduces the theft of intellectual property, personal information, and trade secrets. The use of embedded hardware running on ML-based models is also sustainable as it has a much smaller carbon footprint.

Continuous progress in industry 4.0, including sensors and smart devices that are connected via IoT, makes new applications for AI / ML algorithms. Machines are equipped with hardware and software components for data connection and manipulation. From the point of view of data science, this new model will extract knowledge from monitored assets due to the use of ML and automated improvement methods. One of the most important features of AI / ML in this field is an effective prediction of abnormal behavior in industrial machinery, industrial tools, and processes. In addition, the integration of advanced sensors in artificial intelligence, IoT computational technologies, and cloud computing includes industrial robots, smart decisions, predictive skills, and maintenance with its advanced autonomous behavior in improving production, help, and cooperation with employees and workers in carrying out their duties.

A recent Microsoft study found that American manufacturers using AI performed 11.5% better than those who did not. AI helps industry. Why? The applications are wide. And the statistics are fascinating. According to McKinsey, 50% of companies investing in AI over the next five to seven years will be able to double their cash flow [1].

According to a recent Forbes Insights for AI survey, 44% of automobiles and product developers believe AI will be key to manufacturing over the next five years, while nearly half (49%) said it would be essential to their success.

AI use case

The rapid emergence of AI-based technologies in predictive maintenance takes it to a new level. In the food industry and industrial automation, surveillance ratings have gotten "smarter” than ever in terms of the capabilities they can examine. AI-powered monitoring software can identify a faulty machine before a real problem or unforeseen event occurs. For example, in the factory, the combination of IIoT and technology training allows companies to work very well in the field of various changes such as heat, vibration, light, sound, and humidity of various types of machines.

The technology has been incorporated into asset management for a variety of applications over the years. Given the evolution of technology in the community, it is not surprising that new technologies and applications continue to appear in asset management. Modern technologies, including AI and ML, are based on existing processes and technologies.

We generally divide asset management technology into three main categories:

• User experience and interfaces
• Efficient performance
• Investment activities

In each of these areas, technology helps improve efficiency, risk management, and decision-making. Most importantly, all of these technologies have dedicated individuals and professionals involved in monitoring the results of the technologies and making more informed decisions.

Image and face recognition are classic applications for artificial intelligence and machine learning in production, where they can be used to secure employee access, track attendance and prevent fraud and theft. In another example, the Sony Technology Centre in the UK introduced image processing as part of a project that uses artificial intelligence to monitor product variety and quality on the factory floor. More than 150 high-quality Raspberry Pi SBC and Raspberry Pi cameras were used to evaluate processes such as the assembly of components on separate circuit boards to ensure compatibility.

The Embedded Machine Learning Environment

Industrial and Internet of Things (IoT) applications are now often based on Single Board Computers (SBCs), with some 50 per cent of engineers surveyed by Farnell [2] using these development boards in their designs. Machine learning applications, embedded in a variety of embedded hardware and SBCs, are based on tools and techniques that enable the development and implementation of ML models on resource-constrained nodes. Therefore, the embedded machine learning system comprises hardware vendors, particularly original equipment manufacturers (OEMs), where ML models are deployed and implemented. In addition, it expands the global machine learning ecosystem to include tools and technologies for developing, deploying and implementing ML applications on embedded devices, including IoT devices. In the latter case, ML applications are rightly called AIoT (IoT) devices.

There is now a comprehensive set of embedded hardware that can implement machine learning and deep learning programs. Many devices are inexpensive and can be used flexibly in various IoT applications, some are great for educational purposes. For example, IoT developers familiar with the Arduino ecosystem today can use the Arduino Nano 33 BLE Sense board based on Nordic Semiconductor's SoC to develop TinyML applications. The board has several built-in sensors, including a humidity sensor, temperature sensor, air pressure sensor and microphone, as well as motion sensors, proximity sensors, light colour and light intensity. As such, it is versatile and suitable for a wide variety of uses. Another example is the SparkFuns Edge Development Board, which supports in-depth learning programs like voice transcription and motion detection. The board is based on Ambiq Micros' Apollo3 Blue microcontroller running TensorFlow Lite [3], one of the most popular environments for deep learning applications on embedded devices. Recently, Thunderboard Sense 2, an embedded development platform for developing IoT products, was also upgraded with integrated machine learning functions. The collaboration between SiLabs and Edge Impulse makes it possible to support the development of machine learning applications in various microcontrollers (MCUs), especially MCUs that support EFR32 / EFM32 communication [4]. In this way, diverse business functions can be developed for different applications, such as machine monitoring and audio event analysis.

The development of machine learning models relies on popular data science libraries and tools, such as Python Machine Learning libraries, examples being Scikit-Learn and Keras over Tensorflow, as well as relevant tools such as Jupyter notebooks designed for data scientists and researchers. However, the built-in machine learning ecosystem also includes libraries specially designed to support inferences on devices with limited computing power. This is, for example, the case with TensorFlow Lite, which supports inference on the device. In addition, there are also TinyML [5]-oriented libraries, which allow the deployment of models on devices with very few kilobytes of memory such as microcontrollers. For example, the TensorFlow Lite Micro core can hold 16 KB on an Arm Cortex-M3 and can run various machine learning and deep learning models.

Microsoft is an industry leader in the rollout of its Azure Sphere IoT platform, which also provides the Linux kernel operating system for embedded microcontrollers (MCUs) used in IoT terminals. Common uses include a Starbucks pilot program of integrated MCUs for Azure Sphere to collect telemetry from coffee machines in its stores. The company expects the data it collects to help identify potential problems before repairing coffee machines and other devices in the business. Azure MCUs are inexpensive, offer a variety of connectivity options, including cellular and Ethernet, and can be supported by a variety of development boards and startup kits.

SBCs for AI

Farnell has a wide range of SBC to support embedded AI applications. This includes the most famous SBC, the Raspberry Pi 4 model B which is available with DDR4RAM up to 8 GB.

The Arduino Portenta is a powerful SBC. Asymmetric kernels can run high-level code such as protocol stacks, ML, or even interpreted languages like MicroPython or Javascript at the same time.

Conclusion

In the near future, the convergence of artificial intelligence and embedded systems will lead to enormous advances in image and video recognition. Advances in embedded technology are helping us build imaging devices with higher processing power and smaller footprints. At the same time, AI offers much-needed algorithms for real-time image and video recognition. Implementing these intelligent public safety imaging devices is beneficial because it detects potentially dangerous behavior. Such systems are also used to improve inventory management in factories, control transportation systems, and develop industrial automation.

References: