Resumo:
Fault diagnosis is critical to any maintenance industry, as early fault detection can prevent
catastrophic failures as well as a waste of time and money. In view of these objectives,
vibration analysis in the frequency domain is a mature technique. Although well
established, traditional methods involve a high cost of time and people to identify failures,
causing machine learning methods to grow in recent years. The Machine learning (ML)
methods can be divided into two large learning groups: supervised and unsupervised, with
the main difference between them being whether the dataset is labeled or not. This study
presents a total of four different methods for fault detection and diagnosis. The frequency
analysis of the vibration signal was the first approach employed. This analysis was chosen
to validate the future results of the ML methods. The Gaussian Mixture model (GMM)
was employed for the unsupervised technique. A GMM is a probabilistic model in which
all data points are assumed to be generated by a finite number of Gaussian distributions
with unknown parameters. For supervised learning, the Convolution neural network
(CNN) was used. CNNs are feedforward networks that were inspired by biological pattern
recognition processes. All methods were tested through a series of experiments with real
electric motors. Results showed that all methods can detect and classify the motors in
several induced operation conditions: healthy, unbalanced, mechanical looseness,
misalignment, bent shaft, broken bar, and bearing fault condition. Although all
approaches are able to identify the fault, each technique has benefits and limitations that
make them better for certain types of applications, therefore, a comparison is also made
between the methods.
