In the industrial environment, maintaining a permanent good state of functioning for every piece of equipment has a substantial importance. This, however, is very difficult to attain, due to the mechanical wear, the environment of operation, or improper usage. Predictive maintenance is a practice that is performed to determine the condition of the machinery in service and estimate the time when the maintenance should occur. The challenge of detecting a possible fault in a piece of equipment before it occurs is one of the main tasks of the predictive maintenance process. Reading data from sensors and creating firmware that monitors the equipment can be time and resource-consuming, and not practical if the equipment is changed frequently. Nowadays, the computational power of Artificial Intelligence exceeds that of a computer. As the industrial equipment and the hardware components of a conventional computer are getting increasingly expensive and demanded, more and more entities are running Machine Learning algorithms, which make the data exchange with a server that runs this service a more feasible process. This approach poses several challenges due to latency, privacy, bandwidth, and network connectivity. To solve these limitations, computation should be moved as much as possible towards the Edge, directly on the devices that gather the data. In this article, we propose a compact and low-powered solution that is accurate and small enough to be fitted on a microcontroller or a device that runs on the Edge. This approach ensures that a minimum amount of resources are used. The solution consists of an Unsupervised learning algorithm that can detect anomalies in the vibration patterns of the bearings or the casing of industrial motors. It uses an Autoencoder that takes as input the median absolute deviation of each measurement set provided by an accelerometer, then with the help of a classifier compares the values provided by the output to values that are known to be normal vibration patterns and decides if it deals with an anomaly or not. The low-powered Edge device is an ESP32 board that consumes only 160 mAh on full load but also being powerful enough to maintain WiFi and Bluetooth capabilities when needed. On a more economical operating mode, without WiFi and Bluetooth capabilities it can consume as low as 3 mAh [1]. This feature and the fact that the board is connected directly to the data-gathering sensor makes it preferable to an algorithm hosted on a remote server or a local machine due to low resource consumption and easy maintainability. The Autoencoder is fitted on this board and runs continuously until it encounters an anomaly, which in turn provokes an alert to the user.