<p>Rolling bearing faults are developed gradually from minor faults. The vibration signals of early faults are weak and easily overwhelmed by various interferences and strong noise. The existing methods are always struggled to extract highly discriminative features and the diagnostic performance exhibits weak anti-interference capability. To address these issues, an early bearing fault diagnosis method based on dynamic learning and multi source wavelet kernel net-transformer (MSWKN-T) is proposed in this paper, which includes the dynamic trajectory identification phase, the dynamic feature extraction phase, and the diagnostic model construction phase. In the dynamic trajectory identification phase, the inner dynamic of minor signals in early bearing faults are accurately modeled and identified based on dynamic learning theory. Then, the dynamic trajectories are extracted by embedding the identified information, which can sensitively reflect bearing state changes. In the dynamic feature extraction phase, MSWKN-T is used to extract the crucial semantic information with clear physical significance from the dynamic trajectory. The global interaction ability of local features is realized by calculating the attention weight. In the diagnostic model construction phase, a multi-source wavelet kernel net work is proposed in this paper, which aims to achieve multi-directional and multi-scale feature extraction from dynamic trajectories. It is implemented by expanding multi-direction data source inputs and utilizing multi-head attention mechanism. Compared with the existing methods, the proposed method can extract more discriminative fault features from the complex environment of early bearing faults, and the feasibility and effectiveness have been verified by rigorous testing and validation on the open datasets.</p>