This study addresses the pivotal challenge of fault diagnosis in gears and bearings—crucial components of mechanical systems whose failures can result in unplanned shutdowns, significant economic losses, and safety hazards. Traditional diagnostic methods often falter when faced with the complexity and non-stationarity of vibration signals. To overcome these limitations, we propose a novel TA-YOLO intelligent diagnosis model, leveraging deep attention fusion strategies. Built upon the YOLOv8 framework, our model incorporates the ResNet50 residual module alongside a Triplet Attention mechanism, establishing an innovative “feature extraction-attention enhancement” architecture. The ResNet50 component effectively mitigates the issue of gradient vanishing through its residual connections, while the Triplet Attention mechanism dynamically emphasizes critical fault features across multiple scales. By utilizing Continuous Wavelet Transform (CWT), we convert vibration signals into time-frequency images, enabling end-to-end fault classification. Our experimental results reveal a classification accuracy of 95.56% on the BJTU-RAO dataset, representing a substantial enhancement of 7.41% over the YOLOv8n model. These findings underscore the model’s adaptability and reliability, offering a comprehensive solution for ensuring the safety and operational integrity of mechanical systems in complex environments.