Abstract: As the core equipment of intelligent production lines, the operational status of CNC machine tools not only affects production efficiency but also influences the stability of the entire production line. To address the issues of insufficient diagnostic accuracy and weak generalization capability of fault diagnosis models caused by complex noise interference in real industrial environments, a fault diagnosis model based on time-frequency joint feature analysis and a hybrid attention mechanism is proposed. First, the original vibration signals were processed by a causal denoising and multi-scale time-frequency feature extraction module, effectively suppressing noise interference while preserving multi-scale fault features. Subsequently, a cross-modal interaction network based on a hybrid attention mechanism and score gating was constructed to achieve dynamic weighted fusion of time-domain and frequency-domain features, significantly improving the model's robustness under complex working conditions. The model was validated using a public dataset and a five-axis machining center. Experimental results demonstrate that, under industrial noise interference, the model exhibits significant advantages in feature extraction depth and noise suppression. It provides a reliable algorithmic foundation and technical support for building fault diagnosis systems suitable for complex environments in intelligent production lines.