From Sensor Data to On-Device Inference: An End-to-End TinyML Framework

Sammanfattning

The rapid growth of embedded and Internet of Things (IoT) systems has increased the demand for intelligent data processing directly at the edge. While cloud-based machine learning solutions offer high computational capacity, they often suffer from increased latency, bandwidth usage, and privacy concerns. TinyML addresses these challenges by enabling machine learning inference on resource-constrained microcontroller-based platforms. This thesis explores how an application-specific dataset should be designed, collected, preprocessed, and validated to enable reliable machine learning inference on an embedded edge device. An end-to-end system architecture is proposed and implemented, comprising high-rate accelerometer data collection on an STM32-based platform, a host-side data collection and labeling application, an offline machine learning pipeline, and on-device inference using STM32Cube.AI. Two classification use cases are evaluated. A simple device-position recognition task and a vibration-based anomaly detection task. For each use case, both a feature-based multilayer perceptron (MLP) and a time-domain convolutional neural network (CNN) are trained and evaluated offline. The trained models are subsequently deployed on the embedded platform and evaluated under real-time conditions. Performance is assessed using accuracy, macro- and weighted f1-scores, confusion matrices, inference latency, throughput, and memory footprint. The results demonstrate that reliable real-time inference can be achieved on a microcontroller when data collection and preprocessing are carefully aligned with constraints in the embedded system. The MLP models provide low latency and small memory footprint, making them suitable for simple classification tasks, while CNN models achieve superior performance for vibration-based anomaly detection at the cost of increased computational complexity.