| Rolling bearings and gears are widely used in industrial production as key components in rotating machinery.Due to the working conditions of rolling bearings and gears are complex,the labeled data are scarce for a certain spec component during training;and the vibration signals of components under different specs have large differences,the fault diagnosis model established using a single spec of a certain component is not suitable to directly applied to the fault diagnosis of different specs components.Therefore,it is of great significance to accurately identify the health status of bearings and gears of different specs.Taking deep learning and transfer learning as the core technology,and introducing an emerging contrastive learning framework,a fault diagnosis method based on semi-supervised model contrastive transfer is proposed for rolling bearings and gears to achieve the fault status identification of bearings and gears under different specs.To solve the problem of the lack of labeled samples in the target domain,the efficient diagnostic model cannot be trained,an improved deep model transfer learning method is proposed to transfer the useful knowledge of the source domain to the target domain,and alleviate the lack of data.Short-time Fourier transform is used to obtain the time-frequency spectrum of the vibration signals of different specs components,and image data sets are constructed;the labeled data of a certain specs component are selected as the source domain,and a small amount of the labeled data of other specs components as the target domain.The source domain data are used to train VGG-16 network,SAdam optimizer is introduced to speed up the network convergence speed,and the source domain pre-trained model can be obtained.Then,transfer the pre-trained model parameters of the source domain to the deep network of another spec component to initialize the target domain network parameters,and accelerate the learning and optimization of the network.Due to the large difference between the vibration signals of different specs components,if the traditional model transfer strategy is directly applied to diagnose for different specs,it will easily lead to poor model generalization.In actual,when the target domain contains a very few labels,the direct application of the transfer model will reduce the diagnosis accuracy.For above problems,the proposed method introduce simple framework for contrastive learning of visual representations,and change its projection head activation function to Swish.Learn feature representation by comparing positive and negative samples,the target domain network model is updated adaptively,and the generalization ability of the model after transfer can be improved.Finally,by thawing and fine-tuning a small amount of target domain labeled data,the optimal semi-supervised parameter knowledge transfer fault diagnosis model is established.Experimental results show that the proposed method can achieve the transfer diagnosis between rolling bearings and gears of different specs,and has better stability and higher accuracy. |
