A Comparative Analysis of Quantized TinyML Models vs. Edge AI Frameworks for Real-Time Industrial Anomaly Detection

The deployment of neural network inference at the industrial edge has coalesced around two paradigms: TinyML on microcontroller units (MCUs) designed for kilobyte memory and milliwatt power budgets, alongside Edge AI running on Linux capable singleboard computers (SBCs) where large models can be hosted alongside much higher power demands. Here, we introduce a controlled cross-paradigm benchmark on the task of vibration based industrial anomaly detection within a rotating machinery testbed. A 1D Convolutional Neural Network quantized to INT8 precision is implemented via TensorFlow Lite Micro on three MCU platforms and evaluated against equivalent models running on a Raspberry Pi 4B and NVIDIA Jetson Nano. With the default model size of 280KB, INT8 quantization alone decreases the model storage to that of 72KB while only reduce accuracy by about 1.5 percentage points (94.7?93.2%). The STM32F746ZG can do it with 11ms inference at 0.97mJ per inference, for a energy efficiency advantage of 109× over the Jetson Nano on the same task. Evaluating results on accuracy, latency, memory footprint, power consumption and deployment complexity offers actionable platform selection guidance for constrained industrial IoT scenarios.