Carbon Emissions Prediction and Optimization Method of Hobbing with Multi-source Data Collaborative Based on Federated Learning

Abstract

Sustainability represents a foundational element of Industry 5.0, with low-carbon manufacturing representing a pivotal route to sustainable manufacturing. The Data-driven approach of process parameter optimization for low carbon control is an effective means to reducing carbon emissions during the manufacturing process. However, it relies on large amounts of valid historical data. To overcome the challenge of insufficient valid data within individual enterprises and the necessity for safeguarding data privacy during cross-enterprises collaboration, this paper proposed a method leveraging federated learning to enable multi-source data collaboration across enterprises for training a predictive model of hobbing carbon emissions. Addressing the issue of non- independent identically distribution (non-IID) characteristics inherent in mutil-source hobbing carbon emission data, and the inefficacy of traditional federated learning methods on such complex datasets, this paper designs an improved federated learning algorithm, Federated increment Attentive Message Passing (Fed-I-AMP), to construct the mapping model between hobbing process parameters and carbon emissions. Subsequently, a multi-objective optimization model was developed with the objective of minimizing carbon emissions and Comprehensive Evaluation Index (CEI). This model was solved using the multi-objective Coati algorithm (MOCOA). Experimental validation demonstrates that this approach significantly enhances the convergence performance and prediction accuracy of the model. Additionally, the optimized process parameters provide scientific evidence and practical guidance for low-carbon hobbing operations.