EffiShapeFormer: Shapelet-Based Sensor Time Series Classification with Dual Filtering and Convolutional-Inverted...

Area of Science:

• Machine Learning
• Signal Processing
• Data Science

Background:

• Time series classification is crucial for sensor data analysis in areas like industrial monitoring and fault diagnosis.
• Existing accurate models lack interpretability, while interpretable shapelet-based methods are computationally expensive.
• ShapeFormer, a recent shapelet model, faces challenges with high resource consumption and low training efficiency.

Purpose of the Study:

• To develop an efficient and interpretable framework for sensor time series classification.
• To address the computational and efficiency limitations of existing shapelet-based models.
• To improve both classification performance and model interpretability in sensor data analysis.

Main Methods:

• Proposed Efficiency ShapeFormer (EffiShapeFormer), an efficient framework building on ShapeFormer.
• Introduced a dual-filtering mechanism (Coarse Screening and Class-specific Representation) for efficient shapelet discovery.
• Developed the Convolution-Inverted Attention (CIA) module for synergistic local and global feature extraction.

Main Results:

• EffiShapeFormer demonstrated superior average accuracy and F1-scores across 22 sensor time series datasets.
• Achieved significant improvements in efficiency and performance compared to baseline models.
• Validated the effectiveness of the dual-filtering mechanism and CIA module in feature extraction and classification.

Conclusions:

• EffiShapeFormer offers a significant advancement in efficient and interpretable sensor time series classification.
• The proposed methods effectively balance classification accuracy with computational efficiency.
• EffiShapeFormer presents a promising solution for complex sensor data analysis tasks requiring interpretability.