Graph Neural Network Methodologies For Complex Feature Learning Of Industrial Process Multivariate Data

With the advent of the industrial internet era,modern industry has seen an emergence of diverse data containing high-value information.Effectively learning and utilizing the intrinsic feature knowledge embedded within these data can significantly promote the advancement of traditional industries towards intelligent development.Due to the spatial conduction effects and interactive evolution rules of multiple variables,most multivariate data are located in structured relational systems and exhibit hierarchical interactive relationships.This presents substantial challenges to traditional deep learning methods.Graph neural networks(GNNs),originating from the field of computational social science,provide innovative solutions to these issues.However,further model optimization and transfer research are necessary to adapt GNNs to various industrial application directions.This dissertation aims to conduct an in-depth study of complex feature learning methods based on graph neural networks,and establish effective engineering intelligence models in conjunction with actual applications.The research findings hold significant academic and engineering importance for expanding data-and knowledge-driven intelligent theories and methods,as well as enhancing industrial automation capabilities and levels.Based on the inclination of multivariate data towards expressing different application scenario features,this dissertation defines corresponding representational categories,including temporal causal features,evolutionary difference features,and dependency transformation features.It establishes corresponding graph neural network methods and provides solutions to related engineering problems.The main contributions and innovative achievements of this dissertation are as follows:For temporal causal feature learning,a temporal causal graph attention network(TC-GATN)is proposed to address the hierarchical structured causal relationships in multivariate data.Firstly,a hierarchical directed graph structure is established based on causal analysis theory,ensuring that the direction of information flow aligns with the causal relationships.Secondly,spatial-level gated recurrent units combined with an attention mechanism are introduced within the graph neighborhood.This approach enables nonlinear interactions and adjacent information aggregation of node features through adaptive memory or forgetting,optimizing the original linear aggregation scheme.Finally,a temporal module is supplemented to capture time dependencies.Aiming at the dynamic time-delay characteristics of causal responses,a neural network prediction model based on dynamic time delay reconstruction(DTDRNN)is proposed.Relying on the propagation delay information among variables,the temporal dimension of the driving data is formulated for each modeling target,thereby developing a dynamic mapping pattern between inputs and outputs.This allows the network to accurately capture causal knowledge within multivariate data.The method effectively resolves issues of feature misalignment or information omission caused by static time-delay matching,optimizing the underlying data transmission system.The proposed methods have been applied to predictive modeling for distillation columns and methanol production examples,achieving favorable results.For evolutionary difference feature learning,graph neural networks based on dual-branch learning of multivariate trends and numerical features(TDGNNs)are proposed to address the issue of single data features failing to effectively convey differences between varying operating conditions.In the trend feature learning branch,a trend attention module is designed to aggregate local variation scales endowed with positional encoding,providing novel representational elements to describe differences between process states.Within the parallel graph space of the dual branches,structured interactions and transmissions of trend and numerical nodes are respectively realized.A nodelevel feature fusion layer is embedded to acquire joint information representations.This model benefits from the homologous complementary learning pattern of multivariate data and trends,thereby substantially enhancing the recognition capability of potential evolutionary differences.The proposed method has been applied to fault diagnosis in coal gasification and methanol production,achieving superior identification results.For dependent transformation feature learning,a graph attention autoencoder with symmetric structure(FR-GATE)and a dual-stage feature aggregation graph network model(DA-GNN)are proposed respectively for unsupervised and supervised transformation patterns.In FR-GATE model,an encoding process with forward node information flow and a decoding process with reverse node information flow are designed,constructing a symmetric network structure.This allows node information from the encoding process to be retraced in the decoding process,suppressing excessive node state diffusion.During the reconstruction process,adjacent node states are utilized as additional information sources,maintaining the interaction transmission relationships between variables.The graph space also provides feasibility for regionalized reconstruction of multivariate variables.The method effectiveness is demonstrated through multivariate data reconstruction and anomaly detection tasks in petroleum drilling.For DA-GNN model,during the neighborhood feature aggregation stage,the multivariate time spans are segmented,and subregions of graph nodes are used as the basic aggregation units to reconcile the differences in multivariate interaction scales.In the next stage,a temporal module is introduced to adaptively extract and aggregate subspace states in the temporal dimension,compensating for the inherent lack of temporality in graph models,thus enabling the capture of long-term dependencies.The proposed method effectiveness is validated through soft measurement modeling of chemical oxygen demand in wastewater treatment processes.