As the fourth industrial revolution progressed, data processing and analysis technology is developed. At the same time, machine learning for rotating machine management has been studied using the physical and statistical feature parameters of vibratio...
As the fourth industrial revolution progressed, data processing and analysis technology is developed. At the same time, machine learning for rotating machine management has been studied using the physical and statistical feature parameters of vibration signals.
Machine learning requires a large amount of training data to improve diagnostic accuracy. However, training data is difficult to obtain on the industrial site due to maintenance activities and acquisition of error signals. In addition, current technology is inefficient to classify various conditions(complex fault or fault) of rotating machines.
In order to improve the limitations of these technologies, we reviewed the applicability of the industrial field. And then we found two limitations in this process. First, The formation of training data is performed on the assumption that the normal condition of the rotating machines is almost similar over time. However, many industrial plants perform regular maintenance, and depending on the maintenance results, the normal vibration trend of the equipment is likely to change as well. This reduces the diagnostic performance of machine learning. So, A new training method to minimize the change are developed. The difference signal(delta signal) between fault signal and normal signal is generated with phase synchronization, and training data are formed using an extracted delta signal.
Second, machine learning use genetic algorithms (GA) and principal component analysis (PCA) that in the selection or extraction process of feature conditions of rotating machines. The GA uses only 3 features for displaying result so GA method have a disadvantage with multi-fault classification and also PCA have demerit using all features in any case classify.
In order to overcome these problems, a new method with the advantages of the GA algorithm and the PCA algorithm was proposed to improve the process of machine learning. In the proposed method, appropriate features are selected to classify the machine conditions by the GA, and additional feature selection for performance improvement is repeated. Then the PCA algorithm contains all the information of the features selected from the GA algorithm.
In this paper propose a training method to improve the two limitations of machine learning. The classification performance was evaluated by comparing the proposed method with the original method. As a result, all developed methods have improved performance. The performance of fault diagnosis is more efficient in terms of propose learning than the original method.