Condition monitoring is becoming increasingly important in industry due to the need of increased reliability and decreased loss of production caused by machine breakdown. However, most techniques presently available require a good level of expertise in order to apply them successfully. Therefore, there is a demand for techniques that can make decisions on the running health of a machine automatically and without the need for a specialist to examine the data and diagnose the problem. Artificial neural networks (ANN), fuzzy c-means (FCM), hierchical and partitional clustering, and support vector machines (SVM) have been applied in automated detection and diagnosis of machine conditions. Most of these methods require training based on experimental data both from the healthy as well as the defective machine. This fact significantly diminishes their practical applicability. For this reason, the well established K-means clustering partitional method is proposed in this paper, due to its ease of programming and because it accomplishes a good trade-off between achieved performance and computational complexity. The first step in K-means clustering is finding a set of initial centers. The algorithm is significantly sensitive to the initial randomly selected clusters’ centers. In this paper, in order to overcome this limitation, the initial centers are selected by features extracted from signals simulating the dynamic behavior of bearings under normal condition or under special types of defects (outer or inner race defect). After the selection of the centers of the clusters, the method can be directly implemented to signals measured from the industrial environment. As shown in three different industrial cases, each containing a set of 5 successive measurements following the evolution of a defect, the method is able to identify and appropriately classify the bearing defect. Critical to the success of the method is the type of features extracted from the signals. Some features seem to be irrelevant and redundant in fault diagnosis of rotating machinery, resulting to low diagnosis accuracy. Thus, in order to increase the effectiveness and improve the simplicity of the procedure, a set of features is proposed in this paper, able to clearly characterize the existence and the type of the defect. These features are mostly based on appropriately selected frequency domain parameters. Comparing the proposed frequency domain features to traditional time domain statistical features, indicates the relevance and superiority of the proposed feature set, resulting in 100% success in all 15 measurements of all 3 test cases.

This content is only available via PDF.
You do not currently have access to this content.