Manufacturing Knowledge Graph-Based Disentangled Prediction Framework and Industrial Practices

Quality prediction assumes a pivotal role in manufacturing processes, serving as a linchpin to uphold product uniformity while mitigating the incidence of defects. However, the wealth of prior knowledge inherent in the manufacturing process is often underutilized, leading to limited predictive performance. Simultaneously, the integration of fragmented knowledge in the manufacturing industry often overlooks the ubiquitous presence of time series data, resulting in incomplete knowledge. To address these issues, this paper proposes a manufacturing knowledge graph-based disentangled prediction framework (MFKG-DP Framework). First, a comprehensive manufacturing knowledge graph construction method is proposed, which completes graph relationships through Bayesian network analysis of time series. The second part is the knowledge graph-based disentangled prediction method, which aims to design disentangled mapping matrixes according to the variable relationship provided by the knowledge graph. This facilitates the segmentation and disentanglement of hidden layer features, contributing to enhanced prediction capabilities. Furthermore, this paper incorporates the proposed disentanglement method into deep state space models, creating a disentangled deep state space model called DisSSM. Finally, the effectiveness of the proposed MFKG-DP Framework and DisSSM is validated in industrial processes within manufacturing enterprises.