Monitoring rotating machinery is essential to prevent severe damage, unplanned downtime, and safety hazards. Although machine learning has been widely adopted for fault diagnosis, many existing approaches remain computationally demanding, limiting deployment in resource-constrained environments. In addition, the integration of simulated vibration signals into predictive maintenance workflows remains underexplored. This study proposes a lightweight fault diagnosis framework based on features extracted from the time, frequency, and time-frequency domains of multi-fault vibration signals from rotating machinery. The analysis combines experimental data from the IEEE DataPort COMFAULDA dataset with simulated data generated using Rotordynamic Open-Source Software (ROSS). To enhance discriminative capability at low computational cost, image-based descriptors are applied to time-frequency representations. The models are also deployed on a Raspberry Pi 5 platform. The Support Vector Machine achieved accuracies of 82%, 81%, and 81% using root mean square, power spectral density, and a gray-level histogram extracted from the wavelet-transform representation, respectively. The Random Forest achieved 77%, 57%, and 94% using average mean from envelope, power spectral density, and local binary pattern extracted from the spectrogram, respectively. Overall, the results support lightweight multi-domain machine-learning models for efficient edge-based multi-fault diagnosis in rotating machinery.
Jorge et al. (Fri,) studied this question.