This paper presents an embedded system for diagnosing faults in industrial machines by analyzing vibration signals.The system combines lightweight embedded platforms (ESP32 and NanoPi Neo) with high-resolution IEPE piezoelectric sensors to enable local data acquisition, processing, and classification of mechanical conditions.Machine learning and neural network models are deployed directly at the edge, reducing latency and eliminating the need for cloud-based computation. A full vibration dataset was created for different failure scenarios and made available to the public so that other researchers can use it and reproduce the results. We used and tested a number of classification methods on the NanoPi Neo platform to see how accurate and fast they were. These included Support Vector Machines (SVM), Random Forests, k-Nearest Neighbors (k-NN), and neural networks. The results show that classical machine learning techniques are better than deep learning models in both inference time and classification accuracy when there are embedded limitations. Random Forest had an F1-score of 99.17% with 1.56 ms inference time and SVM obtained 98.99% with 1.33 ms latency. The deep models took a much longer time to compute. The suggested method shows that AI may be used at the edge for real-time issue identification, which is a cost-effective and scalable way to do predictive maintenance that follows the rules of Industry 4.0.
Hakam et al. (Wed,) studied this question.