AstraFormer: A Multiperformance Prediction Method in Manufacturing Processes With Irregular Sampling

An accurate multi-performance prediction is critical for quality control in complex manufacturing processes. However, inherent characteristics such as multivariable coupling, temporal non-uniformity from irregular sampling, and multi-scale dynamics from disturbances pose severe challenges to conventional models. To address these issues, a novel data-driven model named AstraFormer (Adaptive-Scale Temporal Rhythmic Attention Former) is proposed in this paper. The model introduces learnable time interval features to adaptively handle non-uniform time sampling. And a parallel multi-scale convolutional feed-forward network is designed to capture local dynamic patterns at different temporal resolutions. Then, combined with a sparse self-attention mechanism, the model can efficiently model key long-range temporal dependencies. To validate the proposed method, experiments were conducted on a real-world industrial dataset. The results demonstrate that, compared to several baseline models, AstraFormer can achieve a prediction accuracy of over 80% under strict error thresholds. This proves the method can effectively address the challenges of multi-scale and temporal irregularities, providing robust technical support for process optimization and product quality assurance.