Recent years have seen an uptick in the application of transfer learning techniques to improve the flexibility of data-driven approaches, in response to increasing demands from industry for intelligent equipment defect diagnostics with high generalization. There are, however, two important presumptions that are taken as read in the current research: Cross-machine defect diagnostics has adequate labeled data, including fault information, for training models and the feature distribution is consistent across the learning (source domain) and assessment (target domain) data sets. Common issues with distribution alignment, despite its widespread use as a transfer mechanism, include insufficient alignment samples, low confidence in anticipated labels, and misalignment of marginal and conditional distributions. Cross-machine defect diagnosis becomes more difficult due to the bigger domain change between machines. The rotating machine’s simulation model is a useful tool for understanding its features and capabilities, and it can serve as a convenient label for the equipment in a broad scope of conditions. Here, introducing a novel intelligent technique to the problem of cross-machine defect diagnosis by use of an Adaptive Deep Transport Network for Soft Labels (ADTN-SL). The suggested method makes use of a DTN with a broad first-layer subsequent kernel and multiple short convolutional layers to extract features that are portable across machines and to dampen high-frequency noise. Hence, the model named domain adaptation (DA) is employed to mitigate the issues mentioned above. The findings demonstrate that the suggested method outperforms many others in both diagnosis accuracy and transfer performance across a wide range of noise levels and situations. Extensive experiments demonstrate ADTN-SL’s efficacy in cross-machine and cross-location settings.