MCTAN: A Novel Multichannel Temporal Attention-Based Network for Industrial Health Indicator Prediction

Area of Science:

• Industrial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Health indicator prediction is crucial for industrial intelligence.
• Massive multichannel industrial time series data requires efficient processing.
• Existing methods struggle with long-distance temporal relations, channel contributions, and prediction timing.

Purpose of the Study:

• To propose a novel multichannel temporal attention-based network (MCTAN) for industrial health indicator prediction.
• To address challenges in extracting long-distance temporal relations and varying channel contributions.
• To mitigate risks associated with delayed predictions in industrial health monitoring.

Main Methods:

• Developed a multichannel temporal attention-based network (MCTAN).
• Incorporated channel attention to weigh data contributions from different channels.
• Utilized a multi-head local attention mechanism for efficient extraction of long-distance temporal relations.
• Introduced a weighted mean square error loss function to penalize delayed predictions more heavily.

Main Results:

• MCTAN effectively weighs channel contributions and extracts temporal information.
• The multi-head local attention mechanism improves the efficiency of capturing long-distance dependencies.
• The weighted loss function reduces the impact of delayed predictions.
• Experiments on a real-world dataset demonstrated improved prediction accuracy and inference speed.

Conclusions:

• The proposed MCTAN framework offers a systematic approach to industrial health indicator prediction.
• MCTAN enhances prediction accuracy and efficiency by addressing key limitations of existing methods.
• The study provides a valuable tool for industrial intelligence and predictive maintenance.