Key takeaways
AI is transforming manufacturing
XAI reveals why AI makes a prediction
Explanations connect AI with manufacturing expertise
Explainability supports human oversight
The growing use of AI in manufacturing and the need for explainability
AI is becoming increasingly integrated into manufacturing, supporting applications such as automated quality inspection, defect detection, predictive maintenance, machine condition monitoring, and process optimization. Over the past decade, AI and machine learning have developed into important technologies for data-driven manufacturing, enabling automated systems to analyze large volumes of information and support operational tasks. Their growing use, however, also raises a practical challenge: AI-generated results may be difficult to interpret in the context of the physical processes they describe, making human expertise important for assessing whether a prediction is meaningful (Ahangar et al. 2026).
This shift is closely connected to the development of Industry 4.0, which has expanded the use of pervasive sensing, connectivity, and real-time analytics in manufacturing environments. Modern production systems increasingly combine physical equipment with computational infrastructure, generating continuous streams of data on machine states, material variation, and process conditions. These high-dimensional data provide the basis for modelling and monitoring complex manufacturing processes, while deep learning is increasingly used for tasks including defect detection, process modelling, and equipment monitoring across different sensing modalities (Wang et al. 2026).
The growing complexity of these systems makes it important to understand not only what an AI model predicts, but also what information led to that prediction. This is where Explainable AI (XAI) becomes relevant. XAI methods can identify the input features or regions of an image, sound signal, or other data that contributed to a model’s output, providing insight into the basis of an AI-driven decision (Alahmadi, 2026).
Explainable AI in manufacturing
XAI comprises methods designed to provide insight into how AI models generate their predictions and decisions. Rather than treating a model output as a final result, these methods provide additional information about the factors that influenced it, making complex AI systems easier to examine and interpret (Zeiser et al. 2024).
Different manufacturing tasks involve different types of data and model outputs, creating a need for explanation methods that can be adapted to the specific application. For example, an explanation may indicate which image regions contributed to a detected defect or which characteristics of a sensor signal were associated with an abnormal machine state (Ahangar et al. 2026).
The role of XAI is therefore not limited to explaining individual model predictions. It can also provide additional information for examining how AI systems behave across manufacturing applications and whether their outputs are supported by relevant production data. This makes explainability particularly useful when AI is applied to complex processes (Gonzalez, 2024).
Vision-based XAI
Computer vision has become an important component of automated manufacturing inspection, enabling the detection and localization of small surface defects before they develop into more significant quality issues. Deep learning models, including convolutional neural networks and region-based detectors, can achieve high accuracy in recognizing defects, but their predictions may be difficult to interpret. In industrial inspection, this can make it challenging to determine whether a model has identified a genuine defect or has instead relied on irrelevant image characteristics such as background patterns, shadows, or surface variations (Ahangar et al. 2026).
Vision-based Explainable AI addresses this challenge by providing visual representations of the image regions that contribute most strongly to a model’s prediction. Techniques such as Saliency Maps, Gradient-weighted Class Activation Mapping (Grad-CAM), Integrated Gradients (IG), and Occlusion Analysis can highlight influential pixels or image regions, allowing engineers to examine where the model’s attention is concentrated (Ahangar et al. 2026).
Saliency Maps identify pixels that have a strong influence on the prediction, while Grad-CAM provides a broader localisation of the regions contributing to a particular class. Integrated Gradients estimates feature importance by accumulating gradients between a baseline and the input image, whereas Occlusion Analysis evaluates how predictions change when selected image regions are hidden or modified (Ahangar et al. 2026).
These approaches provide complementary ways of examining the behaviour of vision-based models. Their visual explanations can help practitioners verify whether model attention corresponds to actual defect regions and meaningful surface characteristics, providing a practical link between automated image analysis and human inspection expertise (Ahangar et al. 2026).
Explainable AI for time-series and unsupervised models
While vision-based inspection addresses visible surface defects, many manufacturing processes rely on continuous sensor data, such as sound and vibration signals, to monitor equipment condition and detect developing faults. These time-series data are particularly relevant to applications such as predictive maintenance and fault diagnosis, where changes in machine behaviour can provide early indications of abnormal operating conditions. Since labelled fault data are often limited in industrial environments, unsupervised learning methods offer a practical approach by learning patterns of normal operation and identifying deviations from them. Techniques such as Isolation Forests and autoencoders can detect these anomalous patterns without requiring extensive labelled datasets. However, their reliance on unlabelled data can make their predictions difficult to interpret, highlighting the importance of explainability when applying unsupervised models to industrial monitoring (Ahangar et al. 2026).
How explainable AI makes AI-driven decisions more transparent
XAI improves transparency by showing which input features or regions contributed to a model prediction. Depending on the data type, explanations may highlight relevant image regions, signal characteristics, or variables in a multivariate dataset. This allows users to examine the basis of a prediction rather than viewing the model output in isolation (Ahangar et al. 2026).
In manufacturing, these explanations can be compared with known physical and process behaviour. For example, engineers can assess whether features highlighted by an explanation correspond to a machine condition, actual defect, or relevant process change. This provides a basis for evaluating whether model behaviour is consistent with the manufacturing context (Ahangar et al. 2026).
Transparency also supports human-in-the-loop decision-making. The European Commission’s High-Level Expert Group on Artificial Intelligence identifies transparency, human agency and oversight, technical robustness, and accountability as key requirements for trustworthy AI. In manufacturing, explanations can provide additional evidence for interpreting AI outputs, verifying their relevance, and deciding whether further investigation or human intervention is required (Ahangar et al. 2026). XAI methods should therefore be evaluated not only for the consistency of their explanations, but also for their relevance to the physical process, decision context, and needs of the intended user (Madathil et al. 2025).
AI should not only identify problems
As AI becomes increasingly integrated into manufacturing, its role should extend beyond identifying problems to helping people understand and evaluate AI-generated findings. A human-centred approach therefore requires not only accurate predictions, but also meaningful explanations that allow users to assess the basis of automated decisions, particularly when these decisions affect product quality, equipment reliability, or operational safety (Ahangar et al. 2026).
XAI should therefore be viewed not simply as a means of visualizing predictions, but as a tool for connecting AI-based detection with human expertise and diagnostic reasoning. By providing insight into why a problem has been identified, explainability can support human oversight and more informed use of AI in manufacturing (Ahangar et al. 2026).
