Hands‐On Project and Case Studies

Abstract

TinyML is a technology that incorporates artificial intelligence into edge machines, which allows machine learning models to be executed on very low-power microcontrollers with small memory and processing units. In contrast to cloud-based AI-based solutions, TinyML does not require the cloud to run inference on an embedded system, but instead it allows real-time decision-making and is less energy-intensive, less bandwidth-intensive, and less latency-intensive. It can be a good fit in Internet of Things (IoT) systems where power and computational requirements are of significant concern. The chapter is a well-organized, practical introduction to the implementation of TinyML, which includes the choice of the model, its optimization, deployment, and evaluation of the system. It solves the real world issues in implementing deep learning models to predictive maintenance, anomaly detection and environmental monitoring. Inference speed, memory footprint and energy utilization are used to measure performance. Real-world efficiency is measured with profiling instruments like EEMBC(r) MLMarktm and the Power Profiler Kit and trade-offs between accuracy and resource usage are analyzed with them. The discussion proposes both software and hardware ecosystems of TinyML. Arduino supported platforms consist of Nano 33 BLE Sense, STM32 boards and TensorFlow Lite on Microcontrollers-enabled platforms. Model conversion and deployment The TensorFlow Lite, Edge Impulse and TinyMLgen software are used to convert and deploy models. The chapter describes how small convolutional and recurrent neural networks can be applied to either classification or regression to fit into limited hardware. The main case study is concerned with predictive maintenance of industrial motors using the Arduino Nano 33 BLE Sense, which has onboard vibration and temperature sensing. Once accelerator, acoustic, and thermal data has been gathered, post-training quantization and knowledge distillation is used to deploy optimized neural networks. Such methods massively decrease the size of models and their fault detection accuracy remains at more than 90 percent. Transfer learning, pruning, compression using sparsity and integer-only inference are also discussed in order to enhance runtime efficiency. The use of TensorFlow Lite Converter, STM32Cube.AI and GLOW compilers to show deployment workflows is illustrated. The chapter also focuses on hardware-software co-design, whereby microcontroller choice is based on user requirements (in terms of operation), memory, and power requirements. TinyML frameworks with the help of clouds and federated learning methods are explained as they allow updating models safely without violating privacy. Security issues involve encryption and privacy inference protection of the sensitive IoT setting. Multi-edge IoT network scalability is investigated by collaborative and hierarchical learning techniques. The future possibilities are presented as emerging technologies consisting of neuromorphic computing, spiking neural networks and hardware accelerators, as a way of enhancing efficiency. On the whole, this paper will cover the gap between machine learning studies and embedded sensors by offering a practical roadmap towards creating intelligent edge systems requiring energy efficiency, security, and scalability.