Small Sample MKFCNN-LSTM Transfer Learning Fault Diagnosis Method

Abstract

Aiming at the problem that there are all kinds of noise interference in the planetary gearbox of wind turbine in the general experimental scene, the vibration data obtained is less and the fault characteristics are not obvious. A MKFCNN-LSTM migration learning algorithm based on multi-kernel fusion convolution neural network (MKFCNN) and long and short time memory neural network (LSTM) is proposed to realize the fault diagnosis of wind turbine planetary gearbox. Firstly, the MKFCNN is constructed to extract the multi-scale spatial features of the sample signal, and then it is connected in series with LSTM to extract the corresponding time information of the sample signal. In view of the associated fault feature information between the rolling bearing data set and the planetary gearbox data set, the rolling bearing vibration signal of the Western Reserve University is input into the MKFCNN-LSTM as the source domain sample data, and iterative training is used to update the network weight and offset value. The pre-trained MKFCNN-LSTM is obtained, and then fine-tuned by inputting the vibration data of the planetary gearbox with small samples in the target domain, the weight and offset values are transferred from the source domain to the target domain, and finally the accuracy of fault recognition based on the number of small samples in the target domain is improved. The experimental results show that the proposed method can apply the original fault diagnosis knowledge to the vibration data set of the planetary gearbox in the laboratory. Compared with stack autoencoders (SAE), support vector machine (SVM) algorithm, the accuracy of fault identification and classification is improved to a certain extent.