Abstract: (HAtt-Inception) integrated with hybrid attention mechanisms. The model enhances feature learning capability via embedded hybrid attention modules，captures multi-scale feature information through a parallel multi-branch structure，and implements deep training using dual optimizers. Bearing vibration signals are processed by continuous wavelet transform，gramian angular difference field，and markov transition field for feature extraction and dimension conversion. pretraining is performed on the case western reserve university bearing dataset, while a simulated marine operating condition dataset is constructed by superimposing noise interference on the variable-load MFPT dataset. Transfer learning is adopted to verify the modelʹs adaptability under cross-dataset and complex operating conditions， with visualization techniques integrated throughout the research. Tests demonstrate that HAtt-Inception exhibits excellent adaptability and significantly lower detection latency than traditional methods.