Learning Actionable World Models for Industrial Process Control

TL;DR

Addresses active process control under limited data by learning actionable world models that produce interpretable, action-conditioned latent representations. It introduces Enc, P_z, and P_a within a JEPA-inspired architecture, optimized via latent-consistency and action-prediction losses to disentangle action effects in the latent space. Demonstrated on plastic injection molding with small DOE-backed datasets, the approach shows improved action prediction, especially when using a latent predictor and when leveraging transfer learning, while also highlighting risks of negative transfer from unrelated pretraining. The work offers a path toward practical, real-time, robust, data-efficient control in manufacturing settings.

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.

Figures (6)

• Figure 1: Overview of world model training and application for process control. Top: World model training using recorded process signal pairs ($\boldsymbol{x}, \boldsymbol{y}$) that have been generated in the same environment (world) but with changed control parameters (actions) $\boldsymbol{a}=a_1,...,a_n$. The encoder ($Enc$) generates latent representations ($\boldsymbol{z_x},\boldsymbol{z_y}$) used for predicting the actions $\hat{\boldsymbol{a}}$ through the action prediction module ($P_a$). The latent predictive module ($P_z$) guides the learning of the latent representation by incorporating actions, ensuring the representations $\boldsymbol{z}_x$ and $\boldsymbol{z}_y$ align with the process's operational dynamics for more accurate and interpretable predictions. Detailed modeling and training is discussed in Sec. \ref{['method']}. Bottom: Application of the trained world model in real-time process control where the model predicts actions based on a reference process signal plus an observed process signal to maintain stability despite changing external conditions or influences.
• Figure 2: Proposed actionable world model architecture (general case; inputs for injection molding example shown, see Sec. \ref{['sec:application']}). There are $3$ main components in the model, encoder $Enc$, latent predictor $P_z$, and action predictor $P_a$. Losses are computed in representation space and for the action prediction.
• Figure 3: Illustration of the three-dimensional machine parameter space of D1/D2 with a normalized scale. The gray points represent $27$ different machine parameters settings. The green and red points represent the most extreme settings in the corners of the cube (outline indicated by blue lines for clarity), in which individual parameters take on either minimal or maximal values. The pressure curves shown in the dashed rectangles show one example of a reference signal $\boldsymbol{x}$ (green) and two observed signal $\boldsymbol{y}$ (red), each corresponding to different machine parameter settings; the arrows represent an action to transition from $\boldsymbol{x}\xrightarrow{\boldsymbol{a}}\boldsymbol{y}$.
• Figure 4: Visualization of differences between embeddings (either directly as encoded, blue; or as encoded by $Enc$ for $\boldsymbol{x}$ and predicted by $P_z$ for $\boldsymbol{y}$, orange) in the $10$-dimensional embedding space. The red dashed lines indicate a threshold for significantly different activity between the two embeddings. Four different signal pairs are shown: One with a no action (top left), and one each for a change only in mold temperature, injection speed or holding pressure (top right, bottom left and bottom right).
• Figure 5: Visualization of embeddings from D1's test set for Exp. 1 in $2$ dimensions using PCA, colored by the actions that led to the underlying signals. Red dots represent embeddings with no action taken, serving as the reference for the other signals; green, blue and purple dots are created by changing only injection speed, molding temperature or holding pressure.