Nonstationary Discrete Convolution Kernel for Multimodal Process Monitoring.

Area of Science:

• Process monitoring and control
• Machine learning for industrial applications
• Kernel methods in data analysis

Background:

• Kernel transformations enhance data-driven process monitoring, particularly for nonlinear data.
• The radial basis function (RBF) kernel struggles with multimodal process data from various normal operating modes.
• Existing methods for multimodal process monitoring have limitations in accurately describing diverse operational data.

Purpose of the Study:

• To address the limitations of RBF kernels in handling multimodal process data.
• To propose a novel nonstationary discrete convolution kernel for improved process monitoring.
• To enhance fault detection performance in processes with multiple operating modes.

Main Methods:

• Development of a novel nonstationary discrete convolution kernel based on convolution kernel structure.
• Utilizing training samples as the support for the discrete convolution to capture diverse properties.
• Comparison of the proposed kernel against RBF kernels within a kernel principal component analysis (KPCA) framework.
• Evaluation using synthesized examples, numerical simulations, and a benchmark multiphase flow facility dataset.

Main Results:

• The proposed nonstationary discrete convolution kernel effectively describes multimodal process data.
• Demonstrated superior performance compared to RBF kernels in fault detection tasks.
• Validated effectiveness on both simulated and real-world experimental data from a pilot-scale facility.

Conclusions:

• The novel nonstationary discrete convolution kernel offers a significant improvement over traditional RBF kernels for multimodal process monitoring.
• This approach enhances data description and fault detection capabilities in complex industrial settings.
• The proposed kernel provides a robust solution for analyzing processes with diverse operating modes.