Abstract:

Due to the existence of closed-loop feedback system, not all faults will lead to quality deterioration. Quality variables are usually difficult to obtain or have a certain delay. The traditional unsupervised methods cannot judge the impact of faults on quality while detecting whether the process is normal or not. Canonical correlation analysis (CCA), a classical supervised method that can consider the relationship between input and output, has been used for quality-related fault detection. However, process data has problems such as high dimensionality and nonlinearity. The complexity of the system makes CCA more challenging to capture hidden features. This paper proposes a variational automatic encoder-orthogonal canonical correlation analysis (VAE-OCCA) method. First, unsupervised adaptive learning is performed on the input data using a variational automatic encoder to achieve feature extraction for high-dimensional nonlinear process variables. Then, the input-output relationship is considered based on the canonical correlation analysis method, and the obtained correlation coefficient matrix is used to perform singular value decomposition to establish quality-related and quality-independent monitoring statistics. Finally, the effectiveness and superiority of the method proposed in this paper are demonstrated through industrial case tests.

Key words: fault detection, process monitoring, canonical correlation analysis, variational automatic encoder, quality-related

摘要：

由于闭环反馈系统的存在，并不是所有故障均会导致质量发生恶化。质量变量通常难以获得或具有一定的延迟，传统的无监督方法不能在检测过程是否正常的同时判断故障对质量的影响。典型相关分析(canonical correlation analysis，CCA)是一种经典的有监督方法，可以考虑输入输出间的关系，已被用于质量相关故障检测。然而，过程数据存在着维度高、非线性等问题，流程系统的复杂性使得CCA对于隐藏特征的捕获更具挑战性。提出了一种变分自编码器-正交典型相关分析(variational automatic encoder-orthogonal CCA，VAE-OCCA)方法。首先，利用变分自编码器对输入数据进行无监督自适应学习，实现对高维非线性过程变量的特征提取；进而，基于典型相关分析方法考虑输入输出关系，利用得到的相关系数矩阵进行奇异值分解建立质量相关和质量无关监测统计量；最后，通过工业案例测试说明提出方法的有效性及优越性。

关键词: 故障检测, 过程监测, 典型相关分析, 变分自编码器, 质量相关