The Czochralski silicon single crystal growth (Cz-SSCG) process is characterized by strong nonlinear coupling and large thermal inertia, making it difficult for conventional data-driven monitoring methods to effectively capture slow-varying degradation trends. To address this challenge, this paper proposes a novel slow-feature-enhanced dual-stream deep network for intelligent multi-level growth condition monitoring. Unlike traditional single-stream models, the proposed architecture innovatively decouples spatial and temporal dynamics. Specifically, slow feature analysis is employed to extract static and dynamic slow features from multivariate process data, characterizing quasi-stationary operating regimes and long-term evolution trends of the growth process. These decoupled features are processed in parallel by a multi-scale one-dimensional convolutional neural network and a bidirectional long short-term memory network to learn complementary cross-variable and temporal representations. Crucially, a cross-attention fusion module, integrated with a physics-constrained loss function, is designed to adaptively synthesize these representations, effectively suppressing non-stationary noise and aligning predictions with physical process principles. In addition, the SHapley Additive exPlanations method is adopted to quantify the contribution of each input variable, enhancing model interpretability and revealing critical factors that influence crystal growth behavior. Experimental results on real production data demonstrate that the proposed method substantially outperforms conventional approaches in classification accuracy and robustness. Beyond Cz-SSCG, this interpretable and physics-aware framework exhibits strong generalizability, offering a highly adaptable solution for condition monitoring in other complex, multi-phase advanced manufacturing processes.
Ren et al. (Fri,) studied this question.