CSFformer: Redefining multi-channel time series analysis with cross-scale fusion Transformer

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Transformer models with channel independence (CI) excel in time series analysis but struggle with intra-channel noise and long-term trend extraction.
• The fixed receptive field in CI models limits their ability to capture multi-scale temporal features within individual channels.

Purpose of the Study:

• To introduce CSFformer, a cross-scale fusion Transformer designed to overcome the limitations of CI models in multivariate time series analysis.
• To improve the extraction of both short-term fluctuations and long-term trends while handling noise and anomalies.

Main Methods:

• Channel-Independent Masking (CIM) module to refine feature representation by mitigating anomalies and noise.
• Multi-Scale Pyramid Fusion (MSPF) module for extracting fluctuation and trend features across diverse scales.
• Multi-Scale Attention Fusion (MSAF) module to analyze inter-scale interactions and capture complex temporal patterns.

Main Results:

• CSFformer achieved state-of-the-art performance across 7 real-world public datasets.
• Demonstrated superior performance in scenarios with significant fluctuations and trends, such as Traffic and Electricity datasets.
• The proposed modules effectively address noise, anomalies, and multi-scale feature extraction.

Conclusions:

• CSFformer represents a significant advancement in multivariate time series analysis by effectively integrating cross-scale fusion.
• The model's ability to handle complex temporal patterns and noise makes it highly effective for real-world applications.