Intelligent Health Monitoring of Machine Bearings Based on Feature Extraction

Finding reliable condition monitoring solutions for large-scale complex systems is currently a major challenge in industrial research. Since fault diagnosis is directly related to the features of a system, there have been many research studies aimed to develop methods for the selection of the relevant features. Moreover, there are no universal features for a particular application domain such as machine diagnosis. For example, in machine bearing fault diagnosis, these features are often selected by an expert or based on previous experience. Thus, for each bearing machine type, the relevant features must be selected. This paper attempts to solve the problem of relevant features identification by building an automatic fault diagnosis process based on relevant feature selection using a data-driven approach. The proposed approach starts with the extraction of the time-domain features from the input signals. Then, a feature reduction algorithm based on cross-correlation filter is applied to reduce the time and cost of the processing. Unsupervised learning mechanism using K-means++ selects the relevant fault features based on the squared Euclidian distance between different health states. Finally, the selected features are used as inputs to a self-organizing map producing our health indicator. The proposed method is tested on roller bearing benchmark datasets.