Scholarlink

<p>In the era of Industry 4.0, achieving interpretable root-cause tracing in complex industrial processes is a critical prerequisite for intelligent process control. However, the inherent process heredity and strong inter-variable coupling in these systems often obscure disturbance propagation pathways, while conventional black-box AI models lack the transparency required for engineering intervention. To address these challenges, this paper introduces Causally-Aligned Latent Diffusion (CALD), an end-to-end framework that integrates deep generative modeling with causal inference to enable process-level tracing and reliable optimization. The framework is organized as a closed-loop process of representation learning, stage-level diagnosis, and counterfactual prescription. An Industrial Stage-Grouped Latent Diffusion Model (ISG-LDM) is first used to construct a stage-aligned latent representation, followed by a Causal Contribution Quantification Module (CCQM) for stage-level attribution, and finally a counterfactual intervention mechanism for process adjustment. Validated on real-world industrial hot rolling data, the CALD framework demonstrates superior generative fidelity and predictive accuracy (R<sup>2</sup> = 0.9435). Crucially, its counterfactual intervention capability is verified by successfully tracing and correcting anomalous samples to target performance ranges (e.g., adjusting a low yield strength sample from 467 to 526&#xa0;MPa). This work presents a novel paradigm for interpretable diagnostic systems, effectively bridging advanced data analytics with practical engineering decision-making in coupled industrial environments.</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

Causally-aligned latent diffusion modeling for interpretable root-cause tracing in industrial systems

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Pengju Xu,
  • Zhiyan Zhang,
  • Ziqing Zhu,
  • Ri Jin,
  • Zhimin Lv,
  • Keming Hu,
  • Yifan Yan,
  • Qian Sun,
  • Hongjie Li

摘要

In the era of Industry 4.0, achieving interpretable root-cause tracing in complex industrial processes is a critical prerequisite for intelligent process control. However, the inherent process heredity and strong inter-variable coupling in these systems often obscure disturbance propagation pathways, while conventional black-box AI models lack the transparency required for engineering intervention. To address these challenges, this paper introduces Causally-Aligned Latent Diffusion (CALD), an end-to-end framework that integrates deep generative modeling with causal inference to enable process-level tracing and reliable optimization. The framework is organized as a closed-loop process of representation learning, stage-level diagnosis, and counterfactual prescription. An Industrial Stage-Grouped Latent Diffusion Model (ISG-LDM) is first used to construct a stage-aligned latent representation, followed by a Causal Contribution Quantification Module (CCQM) for stage-level attribution, and finally a counterfactual intervention mechanism for process adjustment. Validated on real-world industrial hot rolling data, the CALD framework demonstrates superior generative fidelity and predictive accuracy (R2 = 0.9435). Crucially, its counterfactual intervention capability is verified by successfully tracing and correcting anomalous samples to target performance ranges (e.g., adjusting a low yield strength sample from 467 to 526 MPa). This work presents a novel paradigm for interpretable diagnostic systems, effectively bridging advanced data analytics with practical engineering decision-making in coupled industrial environments.