Time dependent neural network models for detecting changes of state in complex processes: applications in earth...

Area of Science:

• Computational Intelligence
• Data Science
• Time Series Analysis

Background:

• Detecting internal state changes in time-dependent processes is challenging, especially with heterogeneous, multivariate time series data containing imprecision and missing values.
• Existing methods may struggle with the complexity and noise inherent in such datasets.

Purpose of the Study:

• To develop and validate a computational intelligence approach for identifying internal state changes in time-dependent processes.
• To demonstrate the method's capability in handling imprecise data and missing values within multivariate time series.

Main Methods:

• Utilized a computational intelligence approach employing neuro-fuzzy neural networks to model time-dependent non-linear autoregressive processes.
• Employed grid and high throughput computing for model mining using neuro-fuzzy networks and genetic algorithms.
• Generated collections of models and prediction functions to analyze process structures.

Main Results:

• The developed approach successfully identified changes in the internal structure of processes, correlating with state alternations (steady, transient, abnormal, unstable).
• Simulation experiments confirmed the method's sensitivity in detecting subtle state changes.
• The approach proved generalizable and effective across diverse datasets.

Conclusions:

• The computational intelligence framework offers a robust solution for detecting internal state changes in complex, noisy time series data.
• The method has significant potential for applications in earth sciences (e.g., paleoclimate data) and astrophysics (e.g., solar data).
• This approach enhances the understanding of dynamic systems by revealing subtle structural and behavioral shifts.