Abstract:

To address the issues of missing data, imbalanced feature utilization, and limitations of single models in chemical process fault diagnosis, a novel fusion diagnosis method was proposed. First, orthogonal non-negative matrix three-factor factorization was employed to effectively fit the process variable data, impute missing entries in the data matrix, and reconstruct comprehensive relationships among production operating conditions. Second, an SVM model and a DRSN model were trained in parallel to capture linear and nonlinear interactive features, thereby achieving preliminary fault diagnosis. Subsequently, an MLP algorithm was applied to establish a novel mapping relationship at both the result and model levels, enabling model fusion for final fault diagnosis and performance evaluation. Finally, extensive experiments and comparative studies were conducted on the Tennessee Eastman Process benchmark dataset to validate the effectiveness of the proposed method. Experimental results demonstrate that under both single and multiple fault scenarios, the proposed fusion method improves accuracy and recall by an average factor of 1.01 and 1.12, respectively, compared to the best comparative algorithm, thereby significantly enhancing the completeness of fault feature representation and the reliability of diagnostic decision-making while providing robust technical support for the safety monitoring of complex chemical processes.

Key words: chemical process, fault diagnosis method integrating shallow and deep learning (FDSD), support vector machine (SVM), deep residual contraction network (DRSN), multi-layer perceptron (MLP)