The field of Edge AI refers to the development of artificial intelligence that has the capability of processing data and run locally on hardware devices, without necessarily connecting them to the internet. Therefore, processes such as data creation can be performed without the need of uploading or downloading data from the Cloud. A main consequence of the above is the reduction of the response time of a system to extract results on a process. This triggers the development of artificial intelligence applications at the edge. Specifically, on the field of predictive maintenance at industrial level, applications of artificial intelligence at the edge can provide operational state recognition for machines in real time. This diploma thesis presents two methodological approaches to detect three states of operation, for a DC motor. These states are named as good, broken, and heavy load. Initially, for both approaches, features are extracted on the audio data of the IDMT-Isa-ELECTRIC-ENGINE dataset, after undergoing the appropriate pre-processing. A different neural network is then trained with CNN approach. Subsequently, the two models are subject to a post-training quantization process and an appropriate conversion and compression process in order to be inserted into Stm32 Discovery Kit IoT node board. After the completion of the implementations, an experimental application shall be carried out using the board to check the performance of the models on the recognition of the three sound states of the engine’s operation, as well as their response in cases of real-time change of the states. In conclusion, the results of the above procedures are presented, and conclusions are drawn on the performance of the models.