<p>In advanced manufacturing, surface roughness is a key indicator of machined workpiece quality and directly affects product performance and manufacturing efficiency. However, conventional roughness measurement methods, such as contact profilometry and optical inspection, often require precise calibration, are sensitive to environmental disturbances, and exhibit limited data acquisition efficiency. These limitations lead to high measurement costs and low inspection efficiency, making it difficult to satisfy the requirements of real-time monitoring in practical production. To address these issues, this study proposes a multi-source signal fusion model that integrates process parameter conditioning, bidirectional cross-attention, and global feature encoding. First, the collected sensor signals are preprocessed, and feature extraction and dimensionality reduction are then performed using principal component analysis. The PCA-reduced current and cutting force features are used as dynamic inputs, while the process parameters are introduced as static conditions, providing high-quality input data for the prediction model. Second, the PCA-reduced current features, cutting force features, and process parameters are uniformly encoded into token sequences. A FiLM-based linear modulation mechanism is then used to adaptively scale and shift these features, thereby improving prediction accuracy under different process parameter conditions. Finally, bidirectional cross-attention is employed to learn the relationship between the current and cutting force features, which are further fused with the process parameters for global feature encoding. The fused features are fed into a shallow neural network for surface roughness regression. The results demonstrate the effectiveness of the proposed model and provide a feasible solution for surface roughness prediction under practical machining conditions.</p>