[Paper Review] A Review of Explainable Artificial Intelligence in Manufacturing

Georgios Sofianidis, Jože M. Rožanec+2 more|arXiv (Cornell University)|Jul 5, 2021

Explainable Artificial Intelligence (XAI)48 references4 citations

TL;DR

This paper reviews Explainable Artificial Intelligence (XAI) techniques in manufacturing, proposing a taxonomy of model-agnostic and model-specific methods to enhance transparency and trust in AI-driven industrial systems. It evaluates XAI through qualitative and quantitative metrics, demonstrates applications in quality control, cost estimation, and cybersecurity, and highlights the critical role of user-centered explanations in enabling human-AI collaboration for Industry 5.0.

ABSTRACT

The implementation of Artificial Intelligence (AI) systems in the manufacturing domain enables higher production efficiency, outstanding performance, and safer operations, leveraging powerful tools such as deep learning and reinforcement learning techniques. Despite the high accuracy of these models, they are mostly considered black boxes: they are unintelligible to the human. Opaqueness affects trust in the system, a factor that is critical in the context of decision-making. We present an overview of Explainable Artificial Intelligence (XAI) techniques as a means of boosting the transparency of models. We analyze different metrics to evaluate these techniques and describe several application scenarios in the manufacturing domain.

Motivation & Objective

• To address the opacity of AI models in manufacturing, which undermines user trust and decision-making reliability.
• To classify and analyze XAI techniques based on explainability source, scope, and model dependency.
• To evaluate XAI methods using both qualitative and quantitative metrics for industrial applicability.
• To demonstrate practical applications of XAI in quality inspection, production planning, cost estimation, and cybersecurity.
• To advocate for user-informed XAI integration, ensuring explanations align with domain experts’ and operators’ needs.

Proposed method

• Proposes a three-dimensional XAI taxonomy based on: (1) explainability source (intrinsic vs. post-hoc), (2) scope (global vs. local), and (3) model dependency (model-agnostic vs. model-specific).
• Reviews and categorizes XAI techniques such as class activation mapping (CAM), gradient-based saliency maps, LRP, integrated gradients, and guided backpropagation for visual explanation in image-based tasks.
• Introduces a framework for context-aware explanations, incorporating user profile, objective, and focus (local/global) to tailor explanations to different stakeholders.
• Applies model-agnostic methods like LIME and SHAP for feature importance and local explanation generation in tabular and image data.
• Integrates feedback mechanisms in XAI applications to enable user validation and iterative improvement of explanations.
• Employs 3D gradient-weighted CAM for visualizing cost-influencing features in CAD models, enabling interpretable manufacturing cost estimation.

Experimental results

Research questions

• RQ1 How can XAI techniques be systematically classified and evaluated in the context of manufacturing AI systems?
• RQ2 What are the most effective XAI methods for enhancing transparency in image-based quality inspection and defect detection?
• RQ3 How do different explanation types (e.g., visual, textual, feature-based) support decision-making across diverse manufacturing roles?
• RQ4 In what ways can XAI improve trust and security in industrial AI, particularly in detecting data poisoning and model vulnerabilities?
• RQ5 How can XAI be integrated into real-world manufacturing workflows to support human-AI collaboration in Industry 5.0?

Key findings

• XAI techniques such as LRP and guided backpropagation produced well-distributed heatmaps that effectively highlighted relevant image regions in defect classification tasks.
• Converting black-box model predictions into decision trees enabled human-readable textual explanations, achieving the highest level of interpretability for domain experts.
• 3D gradient-weighted CAM successfully visualized machining features influencing manufacturing cost, enabling actionable insights for design optimization.
• XAI methods like CAM and contrastive gradient-based saliency maps improved model interpretability in ultrasonic welding quality assessment, distinguishing true positives from false positives.
• In cybersecurity, XAI enabled identification of perturbed data samples and model vulnerabilities, supporting early detection of data poisoning in smart manufacturing systems.
• Application-grounded evaluation through user feedback on local explanations demonstrated that interactive, context-aware explanations enhance user trust and system usability.