Multiple Classifiers and Data Fusion for Robust Diagnosis of Gearbox Mixed Faults
Authors: Jagath Sri Lal Senanayaka; Huynh Van Khang; Kjell G. Robbersmyr
Extended Abstract:
Detection and isolation of single and mixed faults in a gearbox are essential to enhance the system reliability, lifetime, and service availability. This paper proposes a hybrid learning scheme for gearbox mixed fault diagnosis applications, consisting of multilayer perceptron (MLP)- and convolutional neural network (CNN)-based classifiers, and data fusion for enhancing the robustness. The MLP classifier uses domain knowledge features generated by measuring energies from several frequency bands in the vibration spectrum. The CNN algorithm is trained to identify patterns in the spectrograms of vibration signals via continuous Wavelet transform. A data fusion framework is introduced to improve the robustness and accuracy of the learning algorithms so that the proposed diagnosis scheme can work effectively regardless of noises in the measured data. Within the framework, data fusion is used at feature and decision levels. A Naïve Bayes combiner is selected to fuse results of the individual classifiers at the decision level to enhance the reliability of the fault classification. The robustness of the proposed scheme is tested using two types of noises, namely white Gaussian noise, representing sensor noise and mixing the original vibration signal with another signal, standing for a beating situation in the gearbox. Vibration data from an experimental setup with gearbox mixed faults is used to validate the effectiveness of the algorithm as compared with the counterparts, e.g. support vector machine and autoencoder feature learning algorithms. The comparative study shows that accuracies and robustness of the individual MLP and CNN algorithms are better than those of the compared methods and, can be significantly improved using data fusion at the feature level. Furthermore, the robustness of the algorithm is well secured under noises by combining the diagnosis results of individual classifiers at decision-level.
This study is part of a PhD research project carried during 2016-2019 at the University of Agder Norway, funded by Ministry of Education and Research, Norway. An extended work of the proposed concept for online fault diagnosis application will be available in a forthcoming paper of IEEE Transactions on Industrial Informatics titled, ‘Towards Self-Supervised Feature Learning for Online Diagnosis of Multiple Faults in Electric Powertrains’.