Empirical Sensitivity Analysis of Discretization Parameters for Fault Pattern Extraction From Multivariate Time...

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Time series sensor data analysis for system fault detection presents challenges in feature extraction.
• Data discretization is a common technique to simplify complex time series while retaining essential information.
• Selecting optimal discretization parameters is critical for effective fault pattern identification.

Purpose of the Study:

• To develop a systematic procedure for discretizing multivariate time series data for enhanced fault detection.
• To define and analyze fault patterns and discretization problems within sensor data.
• To provide practical guidance on selecting discretization parameters based on empirical analysis.

Main Methods:

• A systematic discretization procedure involving label definition based on signal distribution and variation trends.
• Label specification to time segments to create discretized state vectors representing system states.
• Empirical sensitivity analysis of discretization parameters to identify informative fault patterns.

Main Results:

• Formal definitions of fault patterns and discretization problems were established.
• The relationship between discretization parameters and sensor signal characteristics was investigated.
• Computational results from ten real-world datasets offer practical parameter selection advice.

Conclusions:

• The proposed systematic discretization procedure aids in identifying informative fault patterns from multivariate time series sensor data.
• The study provides crucial insights into the impact of discretization parameters on fault detection accuracy.
• Practical recommendations are offered for selecting appropriate parameters, improving system fault detection capabilities.