Self-Learning Quality Rules for More Adaptive Statistical Process Control

How can manufacturers identify process deviations reliably when production conditions, products and data patterns are continually changing? ATB presented a new approach to this challenge at the International Conference on Innovative Perspectives on Computational Intelligence and Data Science (InnoComp 2025).

The paper, “Towards Self-learning AI-Based Quality Rules for Statistical Process Control,” was written by Marcel Dechert, Fulya Horozal and Sebastian Scholze from ATB together with Marcel Wabo and Ratan Kotipalli from Continental Automotive Technologies GmbH.

Moving beyond manually defined control rules

Statistical Process Control (SPC) is widely used to monitor manufacturing processes and detect deviations that may affect product quality. Conventional SPC methods typically rely on control charts, manually defined thresholds and expert knowledge. These methods remain valuable, but their application becomes more difficult as production systems grow more complex and generate increasingly large and diverse datasets.

Static thresholds may also lose their relevance when products, process parameters or operating conditions change. Updating the associated quality rules manually requires considerable expertise and can limit the scalability of process monitoring across different production environments.

The paper introduces Self-learning AI-based Q-Rules (SLAQ), an AI-supported methodology for generating and refining quality rules used for anomaly detection in production data. The proposed approach combines optimisation methods with reinforcement learning.

In the first stage, optimisation algorithms analyse historical process data to derive initial quality rules. These rules describe conditions under which the system should identify a process observation as potentially anomalous. In the second stage, reinforcement learning incorporates human feedback to refine the rules. This allows the monitoring logic to learn from the assessments of quality specialists and adapt to changing process conditions over time.

Combining data-driven learning with human expertise

The aim is not to remove quality experts from process monitoring. Instead, the methodology provides a structured way to combine automated data analysis with human knowledge.

Expert feedback remains important because an unusual data pattern does not necessarily indicate a quality problem. Changes may result from planned adjustments, new product variants or other valid operating conditions. By incorporating this feedback into the learning process, the SLAQ approach is intended to improve detection performance while preserving the interpretability of explicit quality rules.

This combination is particularly relevant in industrial settings, where users need to understand why a monitoring system has generated an alert. Explicit rules can make the system’s reasoning easier to inspect than anomaly-detection models that produce classifications without a transparent decision basis.

Research within the DIAZI project

The work was carried out as part of the DIAZI research project, funded by the German Federal Ministry for Economic Affairs and Energy under grant number 13IK012A. The collaboration between ATB and Continental Automotive Technologies links research on adaptive AI methods with requirements arising from industrial quality assurance.

The paper was subsequently published in the Springer proceedings Innovative Perspectives on Computational Intelligence and Data Science, part of the Communications in Computer and Information Science series, Volume 2793.

Authors: Marcel Dechert, Fulya Horozal, Sebastian Scholze, Marcel Wabo and Ratan Kotipalli

Publication: Dechert, M., Horozal, F., Scholze, S., Wabo, M. and Kotipalli, R. (2026): “Towards Self-learning AI-Based Quality Rules for Statistical Process Control.” In: Innovative Perspectives on Computational Intelligence and Data Science. InnoComp 2025, pp. 243–254. Springer, Cham.

Read the paper and publication details on SpringerLink.