Prediction of surface roughness in aluminum alloy milling based on dynamic and static data fusion

In advanced manufacturing, surface roughness is a critical metric for evaluating the quality of machined workpieces, which directly impacts product quality and manufacturing efficiency. However, traditional roughness measurement methods, such as contact profilometers and optical inspection technologies, require precise calibration, are susceptible to environmental influences, and have limited data acquisition capabilities. These methods result in high costs, low efficiency, and an inability to meet the demands of real-time monitoring in actual production. To address these challenges, this study proposes a multi-source signal fusion model architecture that integrates parameter conditions, bidirectional cross-attention, and global encoding mechanisms. First, the collected sensor signals undergo preprocessing, feature extraction, and principal component analysis dimensionality reduction. The reduced-dimensional current and cutting force features serve as dynamic inputs, and process parameters serve as static conditions. This approach provides the model with high-quality data, thereby enhancing prediction accuracy. Second, the reduced-dimensional current and cutting force features, along with the process parameters, are uniformly encoded into token sequences. The FiLM linear modulation mechanism is then applied to adaptively scale and translate these features, improving prediction accuracy under varying parameter conditions. Finally, bidirectional cross-attention learns the relationship between the current and the cutting force. This relationship is then fused with the process parameters for encoding. A shallow neural network is used for regression prediction. The results validate the model's effectiveness in predicting surface roughness, offering a novel solution for this application.