Abstract
In recent years, with the advent of machine learning technology, data-driven fault diagnosis has emerged as a prominent research area. However, since the necessity of training and test data has to follow an identical distribution, the conventional data-driven diagnostic approach is not well-suited to engineering diagnosis. To address this shortcoming, the diagnosis method named FOCAL-TFAM-ResNet-LSTM is proposed to solve the problem of inconsistent distributions. This diagnosis method is mainly composed of two aspects. Firstly, the mapping-based deep transfer learning method with a novel adaptive loss function construction method is proposed. Specifically, the feature-oriented correlation alignment layer (FOCAL) metric function can significantly improve the classification accuracy of transfer learning by alignment of nonlinear distributional differences between the source and target domains in a high-dimensional reproducing kernel Hilbert space. Meanwhile, a novel adaptive metric function based on gradient dynamic weight optimization is proposed to balance the weight allocation relationship between domain and diagnosis, which can simultaneously minimize the domain difference and decrease the classification error. Secondary, the time-frequency attention module (TFAM) is innovatively applied in the ResNet-LSTM network to avoid the gradient disappearance among long-term sequence signals training. Finally, the algorithm is validated by using various experiments during the reciprocating pump dataset and the Case Western Reserve University public dataset. Under extremely adverse conditions (e.g., an SNR of - 5 dB), the FOCAL-TFAM-ResNet-LSTM method attains a classification accuracy of 96.83% and a loss of 0.1024 in reciprocating pump dataset. In comparison to conventional domain alignment techniques such as MMD and MK-MMD, it achieves a notable 8.42% gain in accuracy and a 0.3594 reduction in loss.