In this paper, an ultra-low power embedded unsupervised learning smart vibration sensor is proposed for automatic industrial monitoring and fault detection. Using Kmeans algorithm, it is able to detect abnormal vibrations patterns. Architecture of the system is first presented, then embedded processing algorithms composed of feature extraction and kmeans algorithm are detailed, and finally an implementation on a vibrations simulator machine is described. Results show that faults can be detected with an accuracy of XX% using less than 0.15% of average embedded processor resources on a ARM M4F with an average consumption of 80µW. This smart sensor is relevant for Industrial Internet Of Things (IoT) autonomous monitoring applications, having more than one year of battery life using a single CR2032 coin cell.