Class Incremental Learning And Its Applictions To Fault Diagnosis For Rolling Bearings

Rolling bearings are key components of large machinery because their health states greatly impact the safety of the equipment,and their failure can cause equipment breakdowns,consequently resulting in huge economic losses and even physical casualties.Unfortunately,due to the harsh working environment and other uncontrollable factors,the failure rate of bearings is usually very high.Therefore,studying the fault diagnosis of bearings is of great engineering and academic significance for ensuring equipment performance and preventing fatal accidents.With the development of computer hardware and software technology,bearing fault diagnosis has become increasingly intelligent based on artificial intelligence technology.However,in both traditional intelligent diagnosis methods and deep learning based intelligent diagnosis methods,the batch learning idea is basically adopted,which wastes a lot of time and computing resources.Considering this challenge,this paper proposes a new framework of class incremental learning method based on one-class classifier and classifier combination.Moreover,based on the framework,some research progresses in the intelligent fault diagnosis of rolling bearings under constant speed and variable speed has been made.The main contents and innovations are summarized as follows:(1)To achieve the intelligent fault diagnosis of rolling bearings under constant speed,this paper proposes two fault diagnosis methods based on class incremental learning.The first method is based on support vector data descriptions and needs to extract features manually.In addition,a new support function derived from the distances between the training samples and the center of the hypersphere is proposed.The support function can be applied to combine the support vector data description models of multiple classes for bearings' multi-fault classification.The second method is based on denoising autoencoder and can learn features automatically.A new support function based on the reconstruction error is proposed.The support function can be used to combine the denoising autoencoders of multiple classes for the multi-fault classification of rolling bearings.Experimental results on data of eleven fault types of wheelset bearings in high speed trains verify that the proposed methods have advantages in diagnostic accuracy.(2)To solve the problem of intelligent fault diagnosis of rolling bearings under variable speed,this paper also proposes two fault diagnosis methods based on class incremental learning.These two methods are based on the two methods described in(1),in which the data preprocessing steps are modified.For the first method,two steps,namely,standardization and discrete wavelet transform based signal decomposition,are added in the data preprocessing processes,so that the bearing vibration signal characteristics at variable speed can be extracted.For the second method,in order to eliminate the influence of the change of rotation speed on the calculation of reconstruction error,standardization is added in the data preprocessing processes.Experimental results on data of four fault types collected by machinery fault simulator show that the proposed methods have advantages in diagnostic accuracy.