Broken Rotor Bar Fault Detection and Severity Analysis using Machine Learning Techniques

After training, validating and testing the models available in the Classification Learner app, the LSTM model, and the NPR tool model, it was seen that the Cubic Support Vector Machine provided the best test accuracy of 99.8% for 100% load.   The table below shows the accuracies obtained for each of the models for 12.5% of the load. Model Test Accuracy Train Accuracy ROC (AUC) LSTM     93.04%        92.69%        -    SDNN     96.9%        96.6%        -    Quadratic SVM     96.7%        98.5%        Class 1 – 0.9995        Class 2 – 0.9999     ...

The fifteen statistical time features were obtained using the Extended parks vector transform (EPVA) approach. This approach uses the three-phase current signatures and converts it to direct and quadrature components. The transformation from three phase currents to two phase orthogonal components is computed using the equation given below: After obtaining the direct and quadrature components using the equations given above. The modulus was calculated using the d  and q  components.  The Extended Park Vector Transformation (EPVA) method is used to identify broken rotor defects within the dataset of various classes in a more accurate manner. 

 In order to train our model we had to find some features for each of the classes. The following features were calculated to be fed into our classifier and the neural network trainer.  Mean Max value Root mean square (RMS) Square Mean Root (SMR) Standard deviation Variance Shape factor (using RMS) Shape factor (using SMR) Crest factor Latitude factor Impulse factor Skewness Kurtosis Normalized 5th central moment Normalized 6th central moment

 Since we had visualized the data and had also seen the class separability using PCA and ICA, it was time to move on to training the models. We directly fed in the three phase currents into the classification learner app. Three Phase Currents The group used the classification learner app to train the model, however very bad results were seen.  The percentage accuracies were averaging at around 20 to 50%. This isn't what we required!! We did get an accuracy of 100%! But this was due to applying the ICA data. Obviously, ICA data will give us an accuracy of 100% because it already has the classes separated .  This is wrong!! Now we need to move on to calculate something known as "The 15 statistical time features."

The Independent component analysis (ICA) was able to separate each of the 5 classes very well. ICA was conducted for all the different loads as well.  The figures below shows the class separability for our data. The classes are healthy, 1 BRB, 2 BRB, 3 BRB, 4 BRB. Each figure displays a different set of load. These include: 5Nm 10Nm 20Nm 30Nm 40Nm ICA plot for 5Nm load ICA plot for 10Nm load ICA plot for 20Nm load ICA plot for 30Nm load ICA plot for 40Nm load

PCA finds a projection that captures the largest amount of variation in the data. The original data are projected onto a much smaller space, resulting in dimensionality reduction. We find the eigenvectors of the covariance matrix, and these eigenvectors define the new space. PCA was performed on the dataset after removing the transient component of the data of the BRB dataset. The figure above shows the principal components at load of 5Nm The figure above shows the principal components at load of 10Nm The figure above shows the principal components at load of 20Nm. It can be seen that the blue part (healthy data) moves towards the inside of the plot. The figure above shows the principal components at load of 30Nm. The plots are rotating in an anticlockwise direction. The healthy data at the center remains unchanged. The figure above shows the principal components at load of 40Nm. The same can be seen here as in the previous figure.