Learning Actionable World Models for Industrial Process Control

Peng Yan, Ahmed Abdulkadir, Gerrit A. Schatte, Giulia Aguzzi, Joonsu Gha, Nikola Pascher, Matthias Rosenthal, Yunlong Gao, Benjamin F. Grewe, Thilo Stadelmann

Introduction

In modern industrial manufacturing, the ability to monitor processes in real-time is a standard requirement. However, most existing systems are "passive," functioning primarily as anomaly detectors that alert operators when a process drifts outside of established norms. While identifying a deviation is useful, these systems rarely provide the "why" or the "how to fix it." This paper introduces a novel "actionable world model" designed to bridge the gap between passive monitoring and active control, specifically within the context of complex processes like plastic injection molding.

The Challenge: From Anomaly Detection to Active Control

Industrial processes are characterized by high complexity and often lack precise analytical or simulation models. Traditional machine learning approaches typically produce "black box" representations where various physical factors—such as temperature, pressure, and speed—are inextricably tangled. To achieve fine-grained control, an AI must function as a "digital twin" that understands the causal relationship between specific control actions and the resulting sensor signals. The goal is to move beyond simple detection toward a system that can accurately predict which specific machine adjustments are required to return a deviating process to its ideal state.

Methodology: Action-Aware JEPA

The researchers propose an action-aware self-supervised learning framework based on the Joint Embedding Predictive Architecture (JEPA). Unlike standard generative models that attempt to reconstruct high-dimensional signals, this architecture focuses on learning meaningful representations in a latent space.

Key Components

  • Dual Encoders: The system utilizes two encoders with shared weights to map both a "reference" signal (representing the ideal process state) and an "observed" signal (the current state) into a shared latent space.
  • Latent Predictor ($P_z$): Rather than using a decoder to reconstruct signals, this module predicts the latent state of the observation based on the reference state and a specific action input.
  • Action Predictor ($P_a$): This module analyzes the difference between two latent states to predict the physical action (e.g., "increase temperature by 5 degrees") that caused the transition.
  • Disentangled Representations: A core strength of this architecture is its ability to encourage disentanglement. The model is designed so that specific dimensions in the latent space correspond to specific physical parameters. For instance, a change in injection speed might only trigger activity in specific latent dimensions, while mold temperature affects others, isolating variables for better interpretability.

By calculating loss in the latent space (latent consistency) rather than in the raw signal space, the model effectively ignores irrelevant noise and focuses on the underlying dynamics of the industrial process.

Experimental Validation

The model was validated using real-world data from plastic injection molding across three distinct datasets:

  • D1 & D2: These datasets contain 270 samples each, generated via a systematic "Design of Experiments" (DOE). They vary three primary parameters: holding pressure, injection speed, and mold temperature.
  • D3: A larger, noisier "in the wild" dataset consisting of 460 samples. This dataset involves six parameters, including hot runner temperature and dosing speed, and was used primarily to test the model's transfer learning capabilities.

Key Results and Insights

High Precision and Actionability

The model demonstrated significant accuracy in identifying corrective actions. In the best-performing setup, the model achieved an average action direction error (θ) of only 8.53°. This indicates that the model is highly proficient at determining which "knob" to turn and the magnitude of the required adjustment.

Sample Efficiency

Industrial experiments are expensive and time-consuming. The proposed model remains effective even when training data is reduced by 50%, making it a viable solution for real-world factory settings where data acquisition is a bottleneck.

Transfer Learning Nuances

The research highlighted the potential for pre-training. Using the noisy, larger D3 dataset to pre-train the model significantly improved performance on D1. However, the authors noted instances of "negative transfer" when the source and target domains were too dissimilar, suggesting that the benefits of pre-training depend heavily on domain alignment.

Limitations and Future Work

While the results are promising, several challenges remain. One significant hurdle is the existence of non-bijective relationships, where different combinations of machine parameters can produce nearly identical sensor signals. Resolving this ambiguity is a priority for future iterations. Additionally, the method currently relies on the availability of specific machine parameters during training; if these are not recorded, the model cannot be trained. Future research will focus on scaling this framework to handle more complex processes involving dozens of interacting variables.

Summary updated

Abstract

To go from (passive) process monitoring to active process control, an effective AI system must learn about the behavior of the complex system from very limited training data, forming an ad-hoc “digital twin” with respect to process inputs and outputs that captures the consequences of actions on the process's “world”. We propose a novel methodology based on learning world models that disentangles process parameters in the learned latent representation, allowing for fine-grained control. Representation learning is driven by the latent factors influencing the processes through contrastive learning within a joint embedding predictive architecture. This makes changes in representations predictable from changes in inputs and vice versa, facilitating interpretability of key factors responsible for process variations, paving the way for effective control actions to keep the process within operational bounds. The effectiveness of our method is validated on the example of plastic injection molding, demonstrating practical relevance in proposing specific control actions for a notoriously unstable process.
BibTeX
@inproceedings{yan2025learning,
  author = {Peng Yan and Ahmed Abdulkadir and Gerrit A. Schatte and Giulia Aguzzi and Joonsu Gha and Nikola Pascher and Matthias Rosenthal and Yunlong Gao and Benjamin F. Grewe and Thilo Stadelmann},
  title = {{Learning Actionable World Models for Industrial Process Control}},
  booktitle = {SDS},
  year = {2025},
  url = {https://openreview.net/forum?id=sxQ9jHeyKz},
}
Figure