Scholarlink

<p>Accurate real-time prediction of surface roughness (S<sub>a</sub>) and print time (PT) is essential for ensuring the morphological integrity and production efficiency of Fused Filament Fabrication (FFF) components, particularly in high-precision microfluidic applications. Traditional quality assurance predominantly relies on post-process inspection and static statistical models, which lack the adaptability required for autonomous in-situ control. This study proposes a novel intelligent virtual sensing framework utilizing a hybrid approach that integrates statistical benchmarking with deep learning. A customized FFF system was employed to fabricate 30 specimens based on a Central Composite Design (CCD), systematically varying layer thickness (LT), print speed (PS), material flow rate (MFR), and raster angle (RA). To circumvent the “small-data” limitation of additive manufacturing (AM), a statistically grounded Monte Carlo simulation was utilized to generate 1,000 synthetic samples, leveraging residual error variance from physical ANOVA to capture stochastic process noise. Comparative benchmarking established that a tapered Multi-Layer Perceptron architecture with 5-fold cross-validation (MLP-CV) functioned as the superior virtual sensor, achieving a coefficient of determination (R<sup>2</sup>) of 94.75% for S<sub>a</sub> and 92.50% for PT. Optimal process conditions were identified at LT = 0.15&#xa0;mm, PS = 40&#xa0;mm/s, MFR = 98%, and RA = 30°, yielding a minimum S<sub>a</sub> of 5.85&#xa0;μm. The framework’s practical efficacy was validated through a microfluidic channel case study, where optimized parameters achieved a sub-micrometer precision of 5.35&#xa0;μm. Model interpretability was further verified using SHAP and LIME, confirming that the predictive logic aligns with physical manufacturing principles. With an inference latency of 0.15 millisecond, this framework enables a transition from reactive inspection to proactive, real-time quality assurance, facilitating the autonomous production of complex biocompatible structures essential for advanced healthcare diagnostics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intelligent virtual sensors for real-time quality prediction in FFF 3D printing of microfluidic devices

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • Mahboob Durai M. A.,
  • Denis Ashok S.

摘要

Accurate real-time prediction of surface roughness (Sa) and print time (PT) is essential for ensuring the morphological integrity and production efficiency of Fused Filament Fabrication (FFF) components, particularly in high-precision microfluidic applications. Traditional quality assurance predominantly relies on post-process inspection and static statistical models, which lack the adaptability required for autonomous in-situ control. This study proposes a novel intelligent virtual sensing framework utilizing a hybrid approach that integrates statistical benchmarking with deep learning. A customized FFF system was employed to fabricate 30 specimens based on a Central Composite Design (CCD), systematically varying layer thickness (LT), print speed (PS), material flow rate (MFR), and raster angle (RA). To circumvent the “small-data” limitation of additive manufacturing (AM), a statistically grounded Monte Carlo simulation was utilized to generate 1,000 synthetic samples, leveraging residual error variance from physical ANOVA to capture stochastic process noise. Comparative benchmarking established that a tapered Multi-Layer Perceptron architecture with 5-fold cross-validation (MLP-CV) functioned as the superior virtual sensor, achieving a coefficient of determination (R2) of 94.75% for Sa and 92.50% for PT. Optimal process conditions were identified at LT = 0.15 mm, PS = 40 mm/s, MFR = 98%, and RA = 30°, yielding a minimum Sa of 5.85 μm. The framework’s practical efficacy was validated through a microfluidic channel case study, where optimized parameters achieved a sub-micrometer precision of 5.35 μm. Model interpretability was further verified using SHAP and LIME, confirming that the predictive logic aligns with physical manufacturing principles. With an inference latency of 0.15 millisecond, this framework enables a transition from reactive inspection to proactive, real-time quality assurance, facilitating the autonomous production of complex biocompatible structures essential for advanced healthcare diagnostics.