A Hierarchical Matrix Factorization-Based Method for Intelligent Industrial Fault Diagnosis.

Area of Science:

• Industrial Safety Management
• Machine Learning
• Data Mining

Background:

• Data-driven fault diagnosis is crucial for intelligent industry safety.
• Industrial data often involves mixed physical attributes, challenging traditional shallow models.
• Existing models struggle to leverage coherent information for enhanced diagnostic performance.

Purpose of the Study:

• To present a novel Hierarchical Matrix Factorization (HMF) model for efficient industrial data representation.
• To enhance fault diagnosis capabilities by learning implicit characteristics and high-level features.
• To address nonlinearities in industrial processes with a nonlinear extension (NHMF).

Main Methods:

• Hierarchical Matrix Factorization (HMF) decomposes data through successive matrix factoring into multiple hierarchies.
• Intermediate hierarchies act as analysis operators, learning implicit data characteristics.
• A nonlinear extension (NHMF) incorporates activation functions to handle process nonlinearities.

Main Results:

• HMF and NHMF demonstrate competitive performance in fault diagnosis compared to shallow and deep models.
• The proposed models achieve this performance with reduced computing time compared to deep learning approaches.
• Experimental evaluation on a multiple-phase flow process validates the effectiveness of HMF and NHMF.

Conclusions:

• HMF provides an efficient data representation for improved fault diagnosis.
• NHMF effectively handles nonlinearities, expanding the applicability of the method.
• The proposed models offer a computationally efficient and high-performing solution for intelligent industrial safety.