Efficient gear fault feature selection based on moth‑flame optimisation in discrete wavelet packet analysis domain
Rotating machinery—a crucial component in modern industry, requires vigilant monitoring such that any potential malfunction of its electromechanical systems can be detected prior to a fatal breakdown. However, identifying faulty signals from a defective rotating machinery is challenging due to com...
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2019
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/2301/ http://eprints.uthm.edu.my/2301/1/AJ%202019%20%2810%29.pdf |
| Summary: | Rotating machinery—a crucial component in modern industry, requires vigilant monitoring such that any potential malfunction
of its electromechanical systems can be detected prior to a fatal breakdown. However, identifying faulty signals from a
defective rotating machinery is challenging due to complex dynamical behaviour. Therefore, the search for features which best
describe the characteristic of different fault conditions is often crucial for condition monitoring of rotating machinery. For
this purpose, this study used the intensification and diversification properties of the recently proposed moth-flame optimisation
(MFO) algorithm and utilised the algorithm in the proposed feature selection scheme. The proposed method consisted
of three parts. First, the vibration signals of gear with different fault conditions were decomposed by a fourth-level discrete
wavelet packet transform, and the statistical features at all constructed nodes were derived. Second, the MFO algorithm was
utilised to select the optimal discriminative features. Lastly, the MFO-selected features were used as the input for a support
vector machine (SVM) diagnostic model to identify fault patterns. To further demonstrate the superiority of the proposed
method, other feature selection approaches were applied, including randomly selected features and complete features, and
other diagnostic models, namely the multilayer perceptron neural network and k-nearest neighbour. Comparative experiments
demonstrated that SVM with the MFO-selected features outperformed the others, with the classification accuracy of
99.60%, thus validating its effectiveness. |
|---|