<p>Micro-stops remain one of the most pervasive and least visible sources of performance loss in manufacturing systems, particularly in high-speed, high-variability environments. Their short duration and inconsistent registration hinder Root-Cause Analysis (RCA) and delay Continuous Improvement (CI) cycles. This paper presents a non-intrusive, low-cost edge-IIoT architecture and prototype that combines vibration sensing, autonomous clustering, and compact TinyML models to detect machine states and micro-stops directly on the device. Rather than introducing a new standalone learning algorithm, the work integrates unsupervised pseudo-labelling under real operating conditions with lightweight supervised models for real-time inference in settings where labelled data and controller access are limited. The solution was validated in three complementary contexts: a controlled CNC environment and two industrial cases (a mould machining company and a high-volume beverage production line). Results show that the system accurately distinguishes machine states, achieving over 97.7% accuracy in controlled CNC tests, more than 85% in industrial machining operations, and reliable micro-stop detection in the beverage line, with extended battery autonomy and component costs below €40. These findings show that non-intrusive edge intelligence can improve the visibility of hidden losses and support subsequent root-cause analysis and improvement activities. The work contributes a practical, scalable architecture for edge-based micro-stop monitoring and provides a foundation for future multimodal and multisite IIoT deployments.</p>