Desenvolvimento de uma solução embarcada para detecção de falhas mecânicas em uma bancada didática utilizando machine learning

Abstract

Induction motors are widely used in industrial environments due to their robustness, simple construction, and high efficiency. However, unexpected failures in these machines can lead to unplanned downtime, disrupt production, and result in high operational costs. With this in mind, the objective of this work was to develop an intelligent embedded system for the automatic detection of anomalies and mechanical faults in motors, focusing on unbalance, misalignment, and possible bearing defects, aiming to enhance predictive maintenance strategies. The proposed system uses inertial sensors integrated into the Arduino Nano 33 BLE Sense microcontroller to collect vibration data directly from an experimental mechanical assembly. The raw data acquired is processed locally for feature extraction in the frequency domain using the Fast Fourier Transform (FFT). These features are then used to train an artificial neural network on the Edge Impulse platform, which allows the generation of an optimized model for embedded execution (TinyML), enabling real-time inference directly on the device without the need for external processing. After the development phase, practical bench tests were conducted to validate the system’s performance. The trained model achieved 100% accuracy in identifying the fault classes (unbalance and misalignment) but showed difficulties in correctly classifying the normal operating condition, resulting in an overall accuracy of 83.33%. Further analysis revealed that this limitation was related to changes in the machine’s vibration pattern over time, likely caused by accumulated mechanical wear in the experimental setup. Several adjustment strategies were tested, including the creation of new models with different architectures and datasets, but no significant improvement was observed. The most effective solution involved selectively replacing the samples of the normal class with data that better represented the current operating condition of the machine while preserving the original network structure. The results confirm the potential of embedded systems with machine learning capabilities for mechanical fault diagnosis, offering a more efficient and accessible approach to predictive maintenance.