Search

Purpose: Fault diagnosis is vital to any maintenance sector since early fault detection can avoid catastrophic failures and also a waste of both time and money. Common defect diagnostic methods take just a few features from the vibration signal, which can lead to a wrong analysis. Deep Learning (DL) is well-known for its ability to extract features from a signal, and a Convolutional Neural Network (CNN) is one of the most successful deep learning approaches. Methods: This paper uses a CNN with Short-time Fourier Transform (STFT), a time-frequency feature map, to extract as much information as possible from vibration signals. To validate the method, an experimental bench was used where it was possible to simulate up to six different faults. A vibration signal in the time domain was recorded to obtain the STFT response. Then, a CNN is trained to diagnose and predict the faults, considering the STFT as the only input. Results: The findings suggest that the proposed method can properly identify the various faults. Conclusion: Since the approach is based on frequency domain analysis, it can be easily replicated for different motors.

The detection and classification of failures in electric motors are critical for ensuring operational integrity and financial sustainability in industrial equipment. Traditional methodologies for failure diagnosis have relied on analyzing electrical and mechanical data signatures across time, frequency, and time–frequency domains, necessitating extensive technical acumen and considerable time investment. To address these limitations, this study integrates artificial intelligence, specifically convolutional neural networks (CNNs), with traditional diagnostic signals, thus reducing the requirement for specialized knowledge and expediting the diagnostic process. Employing pre-processing techniques such as wavelet and short-time Fourier transforms, the proposed AI-based method effectively uncovers latent patterns within vibration data from electric motors under varied operational conditions. Our findings reveal that the method achieves optimal classification accuracy (100%) with both STFT and a combination of STFT and wavelet techniques, and slightly less with wavelet alone (98.5%). Additionally, the investigation into the method's resilience against noise affirms its robustness, with accuracies marginally affected even at high noise levels (up to 60%). This research highlights the efficacy and reliability of combining AI with signal processing techniques for fault diagnosis in electric motors.

Fault diagnosis is critical to any maintenance industry, as early fault detection can prevent catastrophic failures as well as be a waste of time and money. Traditional methods involve a high cost in time and people to identify failures, which has caused machine learning methods to grow in recent years, especially unsupervised learning methods. This paper proposes a model to cluster, identify, and diagnose six different failures in electric motors using only uniaxial acceleration signals. For this, the Gaussian mixture model is combined with the Mahalanobis distance. The proposed method is verified through a series of experiments with real electric motors. The results show that the proposed method is an efficient way to identify failures in electric motors, especially bearing, unbalanced, and mechanical loss failures. The accuracy presented values from 93.4 to 100% with an average accuracy of 97.9% considering all cases, showing that the GMM combined with the Mahalanobis distance was able to understand the pattern involved in the vibration signal of the engines. As the experiments are based on real electric motors and faults, the proposed method can be used for early detection of fault conditions.