A Novel Transformer Framework Enhanced by Target Adaptive Attention for Industrial Virtual Metrology

In modern industry, the utilization of data-driven virtual metrology techniques is economically reliable for predicting critical quality variables, which is crucial for enhancing production efficiency and quality. However, existing models still exhibit limitations in predicting key quality variables in industrial processes. This is primarily due to varying requirements across different industrial tasks, where the lack of guidance on critical quality variables results in reduced prediction accuracy. This paper proposes a novel target adaptive attention (TAA) mechanism that operates as the initial processing layer in the Transformer encoder. The TAA mechanism utilizes key quality variables from the task objectives to guide the extraction of features from auxiliary variables, establishing their interrelationships and assigning adaptive weights to individual features. This architecture enables the mechanism to autonomously identify and capture features that demonstrate strong relevance to the critical quality variables, thereby significantly improving the model’s predictive accuracy. Virtual metrology experiments conducted on the debutanizer column process demonstrate that the proposed method achieves accurate predictions of key quality variables, achieving optimal performance across all evaluation metrics.