Enhancing the Distinguishability of Minor Fluctuations in Time Series Classification Using Graph Representation: The...

Area of Science:

• Industrial Sensor Systems
• Machine Learning
• Data Mining

Background:

• Time series classification (TSC) is crucial for incipient fault diagnosis in industrial sensors.
• Existing TSC methods struggle with minor fluctuations in early fault stages, often prioritizing trends over subtle variations.
• This leads to misclassification as low-amplitude fault signals are overlooked.

Purpose of the Study:

• To develop a novel graph-based time series classification framework, MFSI-TSC.
• To enhance the accuracy of incipient fault diagnosis by focusing on minor fluctuations.
• To create a computationally efficient method suitable for resource-constrained sensor systems.

Main Methods:

• MFSI-TSC extracts the trend component from raw time series data.
• Both raw and trend series are converted into graphs representing their 'visible relationship'.
• Graph subtraction isolates differential information, highlighting minor fluctuations for improved distinguishability.

Main Results:

• MFSI-TSC effectively captures and distinguishes minor fluctuations crucial for fault diagnosis.
• The framework demonstrated superior accuracy compared to ten benchmark methods on real-world and public datasets.
• Optimizations were included to reduce computational complexity.

Conclusions:

• MFSI-TSC offers a robust solution for time series classification, particularly for incipient fault detection.
• Its ability to focus on minor fluctuations enhances diagnostic accuracy in industrial sensor systems.
• The method's computational efficiency makes it suitable for deployment in edge computing environments.