Burn-through point prediction and control based on multi-cycle dynamic spatio-temporal feature extraction

Burn-Through Point (BTP) is a critical state in the sintering process, and maintaining a stable BTP is crucial for ensuring the quality of sintered products. However, the complex mechanistic relationships during the sintering process make it challenging to extract meaningful correlations between data, leading to suboptimal performance of prediction-based control methods. To address this issue, this paper proposed a BTP prediction method based on multi-period dynamic spatio-temporal extraction. Building upon this, a comprehensive fuzzy controller based on historical and future state recognition is introduced to achieve stable BTP. Firstly, a time series alignment method based on multi-cycle partitioning is proposed. The Fast Fourier Transform (FFT) operations is introduced to identify hidden data patterns within the observation sequence. Time series alignment is achieved by weighted time delay through fuzzy curve analysis applied to different data patterns. Temporal features are extracted along the temporal dimension using multi-scale 2D convolution, while the graph learning module generates the graph structure by introducing an attentional mechanism to capture the inter-variable dependencies in the learning window. Next, the spatial feature extraction module uses the outputs of the above two modules as inputs to further capture potential spatial features in the time series. Finally, the comprehensive fuzzy controller, by recognizing historical and future states, provides recommendations for the current sintering process speed, stabilizing the sintering process towards the desired operating states. According to the simulation results on actual datasets, this method not only exhibits high predictive accuracy but also effectively maintains control over BTP within a fluctuation range with a mean square error of 0.0109.