Imbalanced class incremental learning system: A task incremental diagnosis method for imbalanced industrial streaming data

Abstract

In recent years, machine learning has been widely used in various fault diagnosis scenarios. However, existing machine learning algorithms tend to work well in closed static environments and handle a constant number of categories. However, in industrial applications, the monitored data is streaming and new categories are constantly introduced into the diagnostic system. In real industrial scenarios, monitoring data is often characterized by an imbalanced distribution due to the uncertainty of machinery operation. The diagnosis of such imbalanced industrial streaming data is known as the imbalanced class incremental learning problem in industrial applications. A novel imbalanced class incremental learning system (ICILS) is proposed for imbalanced industrial streaming data and applied to intelligent fault diagnosis of mechanical equipment. Specifically, a novel graph convolutional sparse autoencoder is firstly designed for the imbalanced dataset to extract feature information with large inter-class scatter. Next, a classification loss function is designed to enhance the classification decision boundary between majority and minority classes by utilizing the prior distribution information of the imbalanced data. Finally, a novel imbalanced class incremental learning rule is derived to realize new class learning without replaying old class data. ICILS has high accuracy and computational efficiency, while ensuring data privacy and security. The designed experiments for the diagnosis of imbalanced industrial streaming data indicated that the proposed method has significant advantages over other methods and can effectively deal with imbalanced streaming data.