Aiming at the problem of insufficient data and label shortage in equipment health status identification, this paper proposes a domain adaptation transfer learning model combined with Morlet wavelet kernel convolution. The model enhances the fault feature extraction capability through Morlet wavelet kernel convolution and tunable-Q wavelet transform sparse representation. To process the unlabeled target domain data, a pseudo label technique and a label dynamic update mechanism are proposed. Firstly, pseudo labels are assigned to the unlabeled data based on feature metrics evaluation, and accordingly, the target domain data are divided into two parts with and without pseudo-labels. Subsequently, a step-by-step training strategy is adopted: first, the pseudo-labelled data is used to train the model with the source domain data to extract similar fault features, and then the feature alignment training is carried out through the domain adversarial framework using the unlabeled data. The label dynamic update mechanism dynamically adjust the pseudo-label assignments during the training process to ensure that high confidence data are involved in the training. Experiments show that this method effectively improves the fault feature extraction effect and diagnosis performance, which is more advantageous than other transfer methods.
Fault signature extraction of rolling bearings under variable speed conditions is crucial while still a challenge due to the non-periodic features of the fault impulses. A sparse representation method based on multi-objective optimization is proposed to extract the non-periodic fault impulses with high fidelity, in which the sparse representation with the generalized minimax concave regularization based on flexible analytic wavelet transform enhanced is adopted. Moreover, a multi-objective optimization model based on the angular domain correlated kurtosis and the harmonic-to-noise ratio of envelope order spectrum is constructed for adaptive parameters optimization of the sparse representation model, upon which a density estimation strategy is proposed to determine the optimal parameters from the Pareto front originally obtained via NSGA-II algorithm. The non-periodic fault impulses can thus be extracted with the fault signature further identified from the envelope order spectrum. The method is validated by analyzing simulation and experiment signals.