Parallel network using intrinsic component filtering for rotating machinery fault diagnosis

Abstract

Abstract Machine learning is gradually applied to the fault diagnosis system of rotating machinery. However, the fault diagnosis system can only classify and identify the fault types previously trained by the model in the system. If the system is required to identify more types of faults, all the untrained new fault types and previously trained fault types need to be input into the model to retrain. Under the current background of big data, the upgrade time of fault types will be relatively long. To solve this problem, a parallel network model based on intrinsic component filtering (PICF) is proposed, in which each type of sample is trained separately, and then each type of training model is reduced in dimension, and finally the model we need is combined. The fault diagnosis framework based on the PICF is proposed. Firstly, the framework divides the input fault samples into training samples and test samples. Then the training samples are randomly segmented and input into the PICF training model, then the activation function is introduced to activate the training features and test features, and finally the softmax classifier is used for classification. The sparsity of order fault training in parallel network is discussed and the influence of sample segment number and nonlinear activation function on diagnosis is studied. Compared with other deep learning methods, the experiment results of the bearing and gearbox show that the proposed method can not only achieve higher fault classification accuracy under small sample training, but also update the model efficiently without reducing the diagnosis accuracy when increasing fault types.