Process fault detection using basic singular spectrum analysis and Kernel Density Estimation, S. Krishnannair

Singular Spectrum Analysis (SSA) is an effective data-adaptive multimodal tool for monitoring chemical processes. However, monitoring multiscale signals using principal component analysis obtained from the multilevel decomposition of process signals using basic SSA assume the upper control limits of monitoring charts (𝑇"𝑎𝑛𝑑 𝑄) to have a Gaussian distribution. This assumption deteriorates the performance of basic SSA when applied to the monitoring of nonlinear processes. To address the issue in this paper, the kernel density estimation (KDE) was used to estimate upper control limits of monitoring charts for basic SSA based nonlinear process monitoring techniques. The monitoring performance of the resulting basic SSA-KDE approach was then compared with a basic SSA approach of which upper control limits of the monitoring charts were determined based on a Gaussian distribution. Application on nonlinear dynamic simulated process and base metal flotation plant chemical processes showed that the SSA-KDE approach provided better overall performance than the SSA approach with Gaussian assumption-based upper control limits for its monitoring charts. Keywords: Fault detection, singular value decomposition, singular spectrum analysis, kernel density estimation