Hybrid Variable Monitoring Mixture Model for Anomaly Detection in Industrial Processes

Area of Science:

• Industrial Process Monitoring
• Statistical Process Control
• Machine Learning for Anomaly Detection

Background:

• Effective process monitoring is crucial for industrial system reliability.
• Modern processes involve both continuous and binary variables, complicating traditional monitoring.
• Existing hybrid variable monitoring (HVM) models have limitations due to strict distributional assumptions (single Gaussian/Bernoulli).

Purpose of the Study:

• To propose an improved algorithm, the HVM mixture model (HVMMM), for anomaly detection in industrial processes with mixed variable types.
• To overcome the limitations of existing HVM models by relaxing strict distributional assumptions.
• To enhance the accuracy and applicability of process monitoring for complex industrial data.

Main Methods:

• Development of the HVM mixture model (HVMMM) incorporating multiple components, each assuming an HVM.
• Application of the expectation-maximization (EM) algorithm for parameter learning of the multiple components.
• Derivation of detailed mathematical expressions for model parameters.
• Analysis of performance improvements attributed to the multi-component approach.

Main Results:

• The HVMMM demonstrates greater suitability for general situations and more accurate data feature characterization compared to standard HVM.
• Numerical and practical case studies validate the effectiveness and efficiency of the HVMMM.
• The multi-component approach led to a 5.49% increase in fault detection rate in a numerical example.
• A 1.6% reduction in the false alarm rate was observed in a practical case study.

Conclusions:

• The proposed HVMMM effectively addresses the limitations of existing HVM models for industrial process monitoring.
• The multi-component strategy significantly enhances anomaly detection performance, improving both sensitivity and specificity.
• HVMMM offers a more robust and accurate solution for monitoring complex industrial processes with mixed data types.