Research On Gearbox Fault Diagnosis Based On Transfer Learning And Attribute Description

As an integrated component widely used in large-scale mechanical equipment,the gearbox can ensure its healthy body and maintain good operation.It is an important guarantee for the normal operation of mechanical equipment.To meet the needs of China's industrial production,mechanical equipment has been designed to be more precise,complex and intelligent.With the promotion of all kinds of precision mechanical equipment,it becomes more and more difficult to collect sufficient mechanical equipment fault data.In recent years,with the rapid development of data-driven deep learning algorithms,more and more experts and scholars build deep learning models for gearbox fault diagnosis.However,most of the existing deep learning models require sufficient labelled data for model training.In the real industrial process,due to the influence of working conditions,environment,design and other factors,there will be few samples(small samples)or even no available samples(zero samples)for model training for a specific fault type,resulting in the problems of cross working conditions with different feature distribution and insufficient training data.Therefore,how to reasonably collect fault data and monitor the health status of mechanical equipment when samples are scarce has become the top priority in the field of fault diagnosis.Based on the above reasons,aiming at the problem that it is difficult to collect enough fault data for gearbox fault diagnosis in the actual production process,this paper makes the following research.(1)A small sample fault diagnosis model combining LSSA optimization and DBN classification is proposed.Aiming at the problem that it is difficult to adaptively adjust the model super parameters when solving the fault diagnosis task of small samples in cross working conditions,resulting in the poor generalization ability of the model,the model is improved from two aspects: initialization optimization and super parameter optimization.When initializing the optimization,the LSSA optimization model is constructed by combining the Logistic chaotic map with sparrow search algorithm(SSA),and the SSA population is initialized by using the Logistic chaotic map,to ensure the uniform distribution of the population;In the process of super parameter optimization,the DBN structure of the training data is optimized through the LSSA optimization model,the source domain training set is pretrained with the optimal structure DBN,and a small number of target domain samples are added for reverse weight optimization.The characteristic parameters are further fine-tuned through reverse error propagation,and finally the accurate identification of the health status of the gearbox in the target domain in the case of small samples is realized.The experimental comparison results show that LSSADBN method obtains more optimization results with less iterations,has fast convergence speed,and has high accuracy when migrating for different target domains.The research of LSSADBN method has certain application value for gearbox fault diagnosis in the case of small samples.(2)A zero sample fault diagnosis model based on attribute description is proposed.Aiming at the zero sample fault diagnosis problem that there are no available samples for model training due to the precision of machine structure and complex operating environment,a new solution to the fault diagnosis problem in the case of zero samples is proposed from the perspective of transfer learning and combined with the idea of zero-shot learning(ZSL),namely X-CNN model.In this model,the Xception network is used as a feature learning module to extract the features of fault signal time-frequency map;The convolutional neural networks(CNN)is used as the attribute learning module to replace the fault features with good attribute description.The attribute matrix is constructed according to the attribute description of the fault category,and the attribute learner trains the known fault attributes so that it can complete the mapping task from features to attributes;The trained attribute learner is used to map the characteristics of test samples,so as to obtain its attribute estimation matrix.Finally,the diagnosis is completed through the similarity comparison of the attribute matrix.The experimental results show that the X-CNN model achieves an average accuracy of 86.5%in the case of zero samples,which is better than the existing zero sample fault diagnosis methods.It is proved that the X-CNN fault diagnosis model can complete the gearbox fault diagnosis without training with test samples,which opens up a new idea for the fault diagnosis of small samples and even zero samples.