A stage-aware deep-supervised bidirectional long short-term memory framework for quality prediction in injection molding

Injection molding is a multi-stage manufacturing process in which process signals exert stage-dependent effects on final dimensional quality. However, existing data-driven models commonly treat the whole molding cycle as a single global sequence or supervise only the final output, which may obscure stage-specific dynamics and weaken cross-stage quality formation. To address these limitations, this paper proposes a multi-stage deep-supervised quality prediction framework that combines selective-kernel channel attention and bidirectional long short-term memory. Full-cycle high-frequency signals are segmented according to process stages, and each stage is equipped with an independent prediction head supervised by the same dimensional target. Homoscedastic uncertainty weighting adaptively balances stage-wise losses and quantifies stage contributions. The framework integrates within-stage multi-scale feature extraction, channel recalibration, temporal dependency modeling, and cross-stage serial bridging to capture stage-specific process dynamics and propagate quality-related information along the molding sequence. Experiments on a real-world injection-molding dataset of 16,600 production cycles demonstrate mean squared error values of (3.05-4.14)×10 −4 for the proposed method across three dimensional prediction tasks, reducing the error by 2.59%-8.96% compared with state-of-the-art methods. Repeated experiments on dimensional target Size 2 confirm statistically significant improvements over baseline sequence models (p < 0.001), with an inference time of 6.495 ± 0.393 ms per sample. Stage-supervision ablation further reveals that removing packing-, cooling-, and filling-stage supervision increases the average mean squared error by 15.30%, 14.93%, and 12.54%, respectively. These findings indicate that the proposed framework can enhance dimensional quality prediction while providing stage-level evidence for interpreting quality formation in injection molding.