Vibration signals of planetary gearboxes have complex components and time-varying characteristics. As the unstable operation of planetary gearboxes leads to unbalanced data distribution within vibration signals, it is difficult to extract gearbox fault information hidden in a large number of data, so the fault diagnosis of planetary gearboxes under nonstationary conditions is highly challenging. For the past few years, intelligent diagnosis methods have been extensively studied in the fault diagnosis field. However, inappropriate signal representations, inadequate training samples, and data differences increase their difficulty in diagnosing planetary gearbox faults. For the above issues, this paper proposes an intelligent diagnostic framework based on time-frequency features and deep residual joint subclass alignment transfer network (DSATN) for planetary gearbox fault diagnosis under nonstationary conditions. One-dimensional vibration signals are converted into the time-frequency representation through signal processing techniques, to reflect the variation of vibration frequency component within the time-frequency domain with time. During the network training, the deep subclass alignment transfer network evaluates the data distributions between relevant subclasses in source and target tasks by the local maximum mean discrepancy. Also, it utilizes a nonlinear transformation to align the global data distributions between both tasks, thus improving the generalization of the trained model in small sample sets. The proposed method is validated through planetary gearbox experiments and achieves good fault classification in the time-frequency domain of nonstationary vibration signals. Different fault categories of gears and planet bearings are all successfully identified.

Intelligent diagnosis methods based on big data have been extensively applied in the fault diagnosis of rotating equipment such as planetary gearboxes. Most of these methods usually satisfy the condition of independent identically distribution among the training and diagnosis data. However, the data distribution in the actual diagnosis task struggles to satisfy the above conditions due to the lack of fault data, missing label information and the feature differences within different signals, thus increasing the difficulty of cross-condition fault diagnosis in small sample sets. Therefore, we propose a dynamic adversarial balance adaptation method with multi-label information confusion (MLC-DABA) for diagnosing planetary gearbox faults under time-varying conditions. In the signal preprocessing process, we transform the nonstationary timing signals into two-dimensional time-frequency matrices for the feature learning of networks, which avoids the frequency characteristic mess caused by frequency overlap. Moreover, we adopt a balance adaptation algorithm to dynamically evaluate the feature distribution between source and target tasks through the domain labels and category labels, thus establishing a balanced adaptation relationship between the feature distributions of both tasks. This dynamic adversarial training mechanism can tap more domain-invariant feature information to measure the distribution distance among tasks, thus closing the feature differences in different tasks and increasing the generalization of source tasks to the data distribution in target tasks. The proposed method is verified in planetary gearbox experiments. Experimental analysis results indicate that the diagnostic performance of MLC-DABA outperforms other comparison methods in terms of accuracy and training robustness.