Abstract: Chemical processes typically involve high temperatures, high pressures and toxic media, making effective fault detection critical for operational safety and stability. Such processes are often characterized by complex variable interactions, where mixed Gaussian and non-Gaussian data distributions coexist with significant nonlinear correlations. However, existing methods frequently rely on Gaussian distribution assumptions or adopt a unified processing strategy for variables with mixed distributions. To break through these limitations, a dual-subspace fault detection method based on principal polynomial analysis (PPA) and deep support vector data description (DSVDD) is proposed. First, the process variables are divided into Gaussian and non-Gaussian subspaces according to their normality, which is achieved by the Kolmogorov-Smirnov (K-S) test. Subsequently, in the Gaussian subspace, a PPA model is constructed to extract nonlinear features of Gaussian variables via polynomial mapping; meanwhile, in the non-Gaussian subspace, a DSVDD model is built using a deep neural network to capture nonlinear features of non-Gaussian variables. Finally, the detection statistics from the dual subspaces are fused via Bayesian inference to construct a new comprehensive statistic for fault detection. The method is verified through the Tennessee-Eastman (TE) process, comparing the PPA-DSVDD dual-subspace method with PPA, DSVDD, and their corresponding subspace detection results, as well as with principal component analysis (PCA), independent component analysis (ICA), and the variable distribution-based statistical process monitoring (VDSPM) method. The results demonstrate that the proposed method effectively enhances the fault detection capability of complex chemical processes by addressing distribution differences and nonlinearity problems, offering practical significance for the safe operation of chemical processes.

Key words: chemical process, fault detection, Kolmogorov-Smirnov test, principal polynomial analysis, deep support vector data description, Bayesian inference

摘要： 化工过程常涉及高温、高压、有毒介质，高效的故障检测对保障化工系统安全稳定运行至关重要。此类过程的变量往往呈现高斯与非高斯混合分布的特征，且存在复杂的非线性耦合，而现有方法依赖高斯分布假设或对混合分布过程采用统一处理策略。为此，本工作研究一种基于主多项式分析(Principal Polynomial Analysis, PPA)和深度支持向量数据描述(Deep Support Vector Data Description, DSVDD)的双子空间故障检测方法。首先通过Kolmogorov-Smirnov (K-S)检验评估过程变量正态性，将变量划分为高斯子空间和非高斯子空间；然后在高斯子空间建立PPA模型，通过多项式映射提取高斯变量的复杂非线性特征，在非高斯子空间构建DSVDD模型，通过深度神经网络提取非高斯变量的非线性特征；最后，通过贝叶斯推理将双子空间的检测统计量进行融合，构造新的综合统计量进行故障检测。通过田纳西-伊斯曼(Tennessee-Eastman, TE)过程对所提方法进行验证，将PPA?DSVDD双子空间方法与PPA和DSVDD及其对应子空间检测结果进行比较，并与主成分分析、独立成分分析及基于变量分布特征的过程监控方法进行对比分析。结果表明，本工作的方法通过解决分布差异和非线性问题，有效增强了复杂化工过程的故障检测能力，对保障化工过程安全稳定运行具有一定的实际意义。

关键词: 化工过程, 故障检测, K-S检验, 主多项式分析, 深度支持向量数据描述, 贝叶斯推理