116
V. Tra and J.-M. Kim
reason is that characterizing AE signals are highly sensitive to the low-energy acoustic
emissions released by developing cracks in bearings [1, 2]. For this reason, this study
designed a bearing fault diagnostic model with the input of AE signals.
The envelope power spectrum-based analysis is the notable approach that uses
the information of characteristic frequencies located in the envelope spectrum of the
signal. The location of these frequencies is closely related to the defective condition
of a bearing. However, these characteristic frequencies are not only dependent on
the crack types of the bearing but also are heavily affected by the rotational speed
of elements in a bearing. In other words, this approach is not effective when being
applied in the context of variable speed. One ideal alternative, which is not affected
by change of operating speed, is the data-driven approach. The fundamental of this
approach is trying to extract as many feature signatures as possible in transformed
domains. These features serve as input to a classification model to determine defective
status as well as the fault type of the bearing. However, it seems to be impossible to
assure that all extracted features are valuable for diagnostic purposes. Some features
may be relevant and redundant which are likely to deteriorate the classification performance. Therefore, the feature evaluation step is necessary to select the most useful
fault-signatures in a given plentiful candidate pool. Through this step, the dimension
of the feature vector is reduced while the characterizing information regarding the
device’s status is still retained.
Two representative feature-selection methods based on the foundation of component analyses are principal component analysis (PCA) [3] and linear discriminant
analysis (LDA) [4]. The PCA method compresses and transforms the original faultfeatures into principal components that can reflect the symptom of a bearing [5].
However, due to the lack of means to calculate inter-category separability estimation, this approach still faces the problem of conserving the most discriminative
features. In contrast, LDA, with the ability to utilize the information of betweencategory and within-category scatter matrices, has shown a great performance in evaluating and selecting discriminative features among original potential fault-signatures.
However, this approach still exits shortcomings. The penalty graph that represents
inter-category separability is sometimes not able to reflect neighborhood relationships between various categories. As a result, the quality of the selected feature-set
can suffer which deteriorates the model’s classification performance. These limitations have prompted us to propose a new genetic algorithm (GA)-based feature
analysis methodology that incorporates well with the k-nearest neighbors (k-NN)
classifier. Experimental outcomes have shown the outstanding ability of the proposed
scheme in picking up the most discriminatory fault-features in a given original candidate pool, regardless of the distribution of initial fault-signatures. In terms of a
classification module, this study uses the k-nearest neighbors (k-NN) classifier, a
nonparametric method that operates well with a limited training set. Besides, it is
also remarkable with greater classification speed, compared to other state-of-the-art
classification algorithms [6].
Following this section, we describe the proposed methodology of bearing fault
diagnostic model in Sect. 11.2. Section 11.3 illustrates solid experiments to prove
the model’s effectiveness. Conclusions are finally shortened in Sect. 11.4.
V. Tra and J.-M. Kim
reason is that characterizing AE signals are highly sensitive to the low-energy acoustic
emissions released by developing cracks in bearings [1, 2]. For this reason, this study
designed a bearing fault diagnostic model with the input of AE signals.
The envelope power spectrum-based analysis is the notable approach that uses
the information of characteristic frequencies located in the envelope spectrum of the
signal. The location of these frequencies is closely related to the defective condition
of a bearing. However, these characteristic frequencies are not only dependent on
the crack types of the bearing but also are heavily affected by the rotational speed
of elements in a bearing. In other words, this approach is not effective when being
applied in the context of variable speed. One ideal alternative, which is not affected
by change of operating speed, is the data-driven approach. The fundamental of this
approach is trying to extract as many feature signatures as possible in transformed
domains. These features serve as input to a classification model to determine defective
status as well as the fault type of the bearing. However, it seems to be impossible to
assure that all extracted features are valuable for diagnostic purposes. Some features
may be relevant and redundant which are likely to deteriorate the classification performance. Therefore, the feature evaluation step is necessary to select the most useful
fault-signatures in a given plentiful candidate pool. Through this step, the dimension
of the feature vector is reduced while the characterizing information regarding the
device’s status is still retained.
Two representative feature-selection methods based on the foundation of component analyses are principal component analysis (PCA) [3] and linear discriminant
analysis (LDA) [4]. The PCA method compresses and transforms the original faultfeatures into principal components that can reflect the symptom of a bearing [5].
However, due to the lack of means to calculate inter-category separability estimation, this approach still faces the problem of conserving the most discriminative
features. In contrast, LDA, with the ability to utilize the information of betweencategory and within-category scatter matrices, has shown a great performance in evaluating and selecting discriminative features among original potential fault-signatures.
However, this approach still exits shortcomings. The penalty graph that represents
inter-category separability is sometimes not able to reflect neighborhood relationships between various categories. As a result, the quality of the selected feature-set
can suffer which deteriorates the model’s classification performance. These limitations have prompted us to propose a new genetic algorithm (GA)-based feature
analysis methodology that incorporates well with the k-nearest neighbors (k-NN)
classifier. Experimental outcomes have shown the outstanding ability of the proposed
scheme in picking up the most discriminatory fault-features in a given original candidate pool, regardless of the distribution of initial fault-signatures. In terms of a
classification module, this study uses the k-nearest neighbors (k-NN) classifier, a
nonparametric method that operates well with a limited training set. Besides, it is
also remarkable with greater classification speed, compared to other state-of-the-art
classification algorithms [6].
Following this section, we describe the proposed methodology of bearing fault
diagnostic model in Sect. 11.2. Section 11.3 illustrates solid experiments to prove
the model’s effectiveness. Conclusions are finally shortened in Sect. 11.4.