Abstract
Industrial Cyber-Physical Systems (ICPSs) are an integral component of modern manufacturing and industries. By digitizing data throughout product life cycles, Digital Twins (DTs) in ICPSs enable a shift from current industrial infrastructures to intelligent and adaptive infrastructures. Thanks to data process capability, Generative Artificial Intelligence (GenAI) can drive the construction and update of DTs to improve predictive accuracy and prepare for diverse smart manufacturing. However, mechanisms that leverage Industrial Internet of Things (IIoT) devices to share sensing data for DT construction are susceptible to adverse selection problems. In this paper, we first develop a GenAI-driven DT architecture in ICPSs. To address the adverse selection problem caused by information asymmetry, we propose a contract theory model and develop a sustainable diffusion-based soft actor-critic algorithm to identify the optimal feasible contract. Specifically, we leverage dynamic structured pruning techniques to reduce parameter numbers of actor networks, allowing sustainability and efficient implementation of the proposed algorithm. Numerical results demonstrate the effectiveness of the proposed scheme and the algorithm, enabling efficient DT construction and updates to monitor and manage ICPSs.
Introduction
In modern smart manufacturing, Digital Twins (DTs) are essential for predictive maintenance and system monitoring. However, two major hurdles exist in their implementation within Industrial Cyber-Physical Systems (ICPSs). First is the problem of Information Asymmetry: to build an accurate DT, a central server requires high-quality sensing data from Industrial IoT (IIoT) devices. However, these devices incur costs to collect and send data, and because the server cannot verify data quality beforehand, low-quality devices may misrepresent their capabilities to claim higher rewards—a classic economic problem known as adverse selection.
Second is the Sustainability of GenAI: while Generative Diffusion Models (GDMs) are excellent at solving complex optimization problems, they are computationally intensive. Training them in an industrial setting can lead to massive energy consumption and carbon footprints, contradicting the goals of "Green Industry."
GenAI-driven DT Architecture and Methods
The researchers developed a multi-layered solution starting with a GenAI-driven DT architecture. This framework defines how GenAI (such as GANs and Diffusion models) assists in four critical stages: data collection, communication modeling, DT maintenance, and decision-making.
To address the incentive problem, the study employs Contract Theory. By designing mathematical contracts consisting of varying data size and reward pairs, the server ensures that IIoT devices naturally select the items intended for their specific sensing capability. This "self-revealing" mechanism forces devices to reveal their true quality to maximize their own utility.
To find the optimal contract in complex, high-dimensional spaces, the authors proposed a Sustainable Diffusion-based Soft Actor-Critic (SAC) algorithm. The core innovation for sustainability is Dynamic Structured Pruning. During the training of the reinforcement learning agent, the algorithm identifies and removes unimportant neurons in the actor network. This reduces the number of parameters, speeding up the model and drastically lowering energy consumption without sacrificing performance.
Key Results and Insights
Numerical simulations and hardware tracking using CodeCarbon revealed several key insights:
- Superior Performance: The proposed diffusion-based SAC outperformed standard Deep Reinforcement Learning (DRL) algorithms like PPO, standard SAC, and DDPG in terms of total utility for the DT server and sensing devices.
- Massive Emission Reduction: The "sustainable" version of the algorithm (with 10% pruning) produced only 0.38g of CO2, compared to 32.10g for the unpruned version—a reduction of approximately 98.8%.
- Efficiency and Generalization: Even with a 50% pruning rate, the model eventually converged to a high reward. Interestingly, pruning was found to improve model generalization by removing redundant parameters that might otherwise overfit to industrial noise.
Conclusion and Future Work
This work demonstrates that GenAI can be both effective and sustainable in industrial settings. By combining contract theory with pruned diffusion models, the researchers solved the incentive problem while drastically lowering the environmental footprint of AI training. Future research will explore decentralized multi-server environments and the deployment of these mechanisms in live, physical factory floor environments to move beyond high-fidelity simulations.
Summary updated