Encoding Time Series as Multi-Scale Signed Recurrence Plots for Classification Using Fully Convolutional Networks.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Deep neural networks (DNNs) show promise in time series classification (TSC).
• Encoding time series as recurrence plot (RP) images leverages DNNs for improved accuracy.
• Existing RP methods struggle with sequence length variability and tendency confusion.

Purpose of the Study:

• To propose a novel method for time series classification using improved recurrence plots.
• To address limitations in handling sequence length and tendency confusion in time series data.
• To enhance the accuracy and visualization of time series classification.

Main Methods:

• Introduced Multi-scale Signed Recurrence Plots (MS-RP) by incorporating phase space dimension and time delay embedding.
• Developed an asymmetrical structure to represent long sequences (>700 points).
• Utilized sign masks to mitigate tendency confusion and trained a Fully Convolutional Network (FCN) with MS-RP images.

Main Results:

• The proposed MS-RP method, combined with FCN, demonstrated superior performance on 45 benchmark datasets.
• Achieved state-of-the-art classification accuracy.
• Showed improvements in visualization evaluation compared to existing methods.

Conclusions:

• The MS-RP approach effectively handles challenges associated with sequence length and tendency confusion in TSC.
• The novel method offers a significant advancement in time series classification accuracy and interpretability.
• This work provides a robust framework for leveraging deep learning with enhanced recurrence plots.