Dual-Level XAI-Guided Digital Twin Framework for Prescriptive Decision Making in Sensor-Driven Manufacturing

Abstract

Achieving Zero-Defect Manufacturing (ZDM) in precision micro-injection molding requires continuous monitoring of non-linear interactions among continuous thermodynamic and kinetic sensor data (e.g., melt temperature, injection pressure) and discrete equipment states. The high-throughput production of optical lenses operates under stringent physical boundaries, limiting spherical power deviations to a ±0.25 Diopters (D) threshold. This study proposes a sensor-driven digital twin framework for virtual metrology (VM) and prescriptive decision support. The proposed hybrid framework operationalizes Knowledge-Informed Machine Learning (KIML) by imposing physical constraints during evolutionary optimization and auditing the learned internal interactions via a dual-level Explainable AI (XAI) protocol. Utilizing 175,089 sensor logs, a continuous surrogate is constructed via the Feature Tokenizer Transformer (FT-Transformer). Under randomized 5-fold cross-validation, the surrogate achieves a Root Mean Squared Error (RMSE) of 0.4211 D and a Coefficient of Determination (R2) of 0.97, demonstrating a 16.8% error reduction over the 1D-CNN baseline. To rigorously evaluate inter-machine generalization and mitigate batch-level data leakage, an equipment-isolated Leave-One-Machine-Out (LOMO) protocol confirms structural robustness with an R2 of 0.8820. Furthermore, residual distribution analysis demonstrates that 76.19% of the validation samples actively fall within the ±0.25 D physical tolerance. Cross-verifying intrinsic Multi-Head Self-Attention (MHSA) weights against a global proxy statistically captures the underlying associative affinities between hardware and continuous sensor metrics. Integrating this differentiable surrogate with a real-valued Genetic Algorithm (GA) enables the autonomous generation of optimized process recipes, achieving an algorithmic convergence error below 0.001 D within the continuous latent space. While future physical validation remains necessary, this framework establishes a transparent, auditable foundation for prescriptive smart manufacturing.

The authors' abstract, as published at the source. Sensors, 2026 · DOI ↗