Enhancement of bearing defect diagnosis via genetic algorithm optimized feature selection
The main objective of this research is to enhance the classification performance of the neural network-based bearing fault diagnostic module particularly when the input data has unpredictable variations compared to the training data under various working conditions. The most challenging problem in t...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | en en |
| Published: |
2015
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/43794/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/43794/2/FULLTEXT.pdf https://eprints.ums.edu.my/id/eprint/43794/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The main objective of this research is to enhance the classification performance of the neural network-based bearing fault diagnostic module particularly when the input data has unpredictable variations compared to the training data under various working conditions. The most challenging problem in the fault diagnosis tasks is classifying testing data that has never been seen before by the classifier during training. Therefore genetic algorithm (GA) is employed to search for a minimum number of relevant features nonlinearly to increase the classification accuracy while reducing the computational effort of the training process. However this feature selection algorithm might be unstable due to the stochastic property of GA. In addition, GA has the limitation on generalization which causes the problem of overfitting to the training data. Therefore a correlation-based filtering algorithm is embedded into GA feature selection to solve the over-fitting problem and increase the adaptability of the diagnostic scheme to unpredictable input data. The developed bearing fault diagnosis system has been evaluated and assessed for various working conditions such as rotating speeds, bearing types, fault types and fault sizes. Results show that the reinforced network classifier with GA feature selection algorithm has successfully increased the classification accuracy of training process and testing process by 13.87% and 14.21% respectively compared to the conventional neural network classifier. However the average classification accuracy of 84.74% on the unseen test data did not achieve the acceptable average success rate of 90% in this application. This is due to the features selected in this classifier is over-fitted to the training data and not generalized for variations in testing data. Subsequently, the integration of embedded correlation-based filtering algorithm has further increased the classification accuracy of training process and testing process by 4.93% and 14.73% respectively. The average classification accuracy of 99.47% on the test data achieved the acceptable average success rate. Thus, it can be concluded that the developed algorithm is capable to improve the classification efficiency by improving the generality of the classifier in classifying test data with unpredictable variations under various working conditions. |
|---|