Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Artificial Neural Twin -- Process Optimization and Continual Learning in Distributed Process Chains

Johannes Emmert, Ronald Mendez, Houman Mirzaalian Dastjerdi, Christopher Syben, Andreas Maier

TL;DR

The paper tackles decentralized industrial optimization under data sovereignty constraints and distribution drift, proposing the Artificial Neural Twin (ANT) that unifies MPC, Bayesian differentiable data fusion, and decentral sensor networks. ANT enables state inference across distributed process steps and backpropagation of losses to local parameters or AI-models, while preserving privacy by exchanging only material-stream information. The authors develop a differentiable data fusion framework with MAP derivatives, incorporate Covariance Intersection to handle unknown correlations, and implement an ANT algorithm with information and backpropagation phases orchestrated by a loss node. A Unity-based virtual plastic recycling facility demonstrates state estimation accuracy and near-optimal parameter convergence under constant and dynamic mass-flow scenarios, illustrating ANT’s potential for decentralized optimization and continual learning without data-model sharing. The work suggests ANT as a flexible, model-agnostic connector capable of retraining AI-based sensors and process models in response to drift, with practical relevance for industry and data-security concerns.

Abstract

Industrial process optimization and control is crucial to increase economic and ecologic efficiency. However, data sovereignty, differing goals, or the required expert knowledge for implementation impede holistic implementation. Further, the increasing use of data-driven AI-methods in process models and industrial sensory often requires regular fine-tuning to accommodate distribution drifts. We propose the Artificial Neural Twin, which combines concepts from model predictive control, deep learning, and sensor networks to address these issues. Our approach introduces differentiable data fusion to estimate the state of distributed process steps and their dependence on input data. By treating the interconnected process steps as a quasi neural-network, we can backpropagate loss gradients for process optimization or model fine-tuning to process parameters or AI models respectively. The concept is demonstrated on a virtual machine park simulated in Unity, consisting of bulk material processes in plastic recycling.

The Artificial Neural Twin -- Process Optimization and Continual Learning in Distributed Process Chains

TL;DR

The paper tackles decentralized industrial optimization under data sovereignty constraints and distribution drift, proposing the Artificial Neural Twin (ANT) that unifies MPC, Bayesian differentiable data fusion, and decentral sensor networks. ANT enables state inference across distributed process steps and backpropagation of losses to local parameters or AI-models, while preserving privacy by exchanging only material-stream information. The authors develop a differentiable data fusion framework with MAP derivatives, incorporate Covariance Intersection to handle unknown correlations, and implement an ANT algorithm with information and backpropagation phases orchestrated by a loss node. A Unity-based virtual plastic recycling facility demonstrates state estimation accuracy and near-optimal parameter convergence under constant and dynamic mass-flow scenarios, illustrating ANT’s potential for decentralized optimization and continual learning without data-model sharing. The work suggests ANT as a flexible, model-agnostic connector capable of retraining AI-based sensors and process models in response to drift, with practical relevance for industry and data-security concerns.

Abstract

Industrial process optimization and control is crucial to increase economic and ecologic efficiency. However, data sovereignty, differing goals, or the required expert knowledge for implementation impede holistic implementation. Further, the increasing use of data-driven AI-methods in process models and industrial sensory often requires regular fine-tuning to accommodate distribution drifts. We propose the Artificial Neural Twin, which combines concepts from model predictive control, deep learning, and sensor networks to address these issues. Our approach introduces differentiable data fusion to estimate the state of distributed process steps and their dependence on input data. By treating the interconnected process steps as a quasi neural-network, we can backpropagate loss gradients for process optimization or model fine-tuning to process parameters or AI models respectively. The concept is demonstrated on a virtual machine park simulated in Unity, consisting of bulk material processes in plastic recycling.
Paper Structure (29 sections, 31 equations, 11 figures, 3 tables)

This paper contains 29 sections, 31 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Backpropagation
  4. Differentiable Data Fusion
  5. Statistical Data Fusion
  6. Covariance Intersection
  7. MAP derivatives and their use in backpropagation
  8. ANT algorithm
  9. AI-model fine tuning
  10. Results
  11. Virtual facility
  12. Process models
  13. ANT-facility data exchange
  14. Machine performance optimization
  15. Evaluation
  16. ...and 14 more sections

Figures (11)

  • Figure 1: Backpropagation through a two layer network or two process steps. The forward pass (blue) amounts to the decentral data fusion and state estimation. The backward pass utilizes
  • Figure 2: Bayesian inference, for distributed data fusion and prediction.
  • Figure 3: Communication network between time steps and processes, for information exchange and gradient propagation.
  • Figure 4: Machine components: Sieving machine (left), showing the rotational speed (optimization parameter), inlet, and outlets for three object sizes. Magnetic sorter (right), showing the force field, magnet distance (optimization parameter), and the outlets for two material types
  • Figure 5: Flow diagram of the process. Input on sieving machine, afterwards three long conveyor belts and at the end three magnetic sorters, with corresponding flow sensors (Fl) in the respective sections of the process.
  • ...and 6 more figures