Hardware-Aware Real-Time TinyML for Industrial IoT Applications

Edge intelligence and Tiny Machine Learning (TinyML) have become key enablers for real-time, on-device intelligence in industrial IoT applications, such as predictive maintenance, anomaly detection, and process monitoring. TinyML allows inference to be performed locally, reducing latency, enhancing data privacy, and lowering energy consumption compared to cloud-based solutions. Modern heterogeneous edge platforms, equipped with CPUs, GPUs, and deep learning accelerators (DLAs), provide significant computational resources; however, these are often underutilized due to the lack of hardware-aware scheduling strategies. This project focuses on developing a dynamic TinyML framework that efficiently executes lightweight neural network models on heterogeneous accelerators, while incorporating techniques such as model compression, quantization, and adaptive inference. The framework will be evaluated on industrial IoT datasets, including NASA CMAPSS for predictive maintenance and SWaT for anomaly detection, measuring latency, throughput, energy consumption, memory usage, and predictive accuracy to demonstrate practical real-time deployment and efficiency on modern edge devices.