Time Series Forecasting of Motor Bearing Vibration Based on Informer.

Area of Science:

• Engineering
• Data Science

Background:

• Motor bearing vibrations are critical indicators for equipment health and energy efficiency.
• Conventional time series forecasting methods (CNN, LSTM) suffer from error accumulation, especially for longer predictions.

Purpose of the Study:

• To address the error accumulation issue in motor bearing vibration forecasting.
• To introduce and evaluate the Informer model for improved time series prediction accuracy.

Main Methods:

• The study utilizes the Informer model, an architecture based on Transformer, incorporating ProbSparse self-attention and self-attention distilling.
• Model parameters were optimized using random search to enhance forecasting performance and efficiency.
• Informer's performance was benchmarked against other forecasting models on three public datasets.

Main Results:

• The Informer model demonstrated superior performance in time series forecasting of motor bearing vibrations.
• Achieved significant error reduction, with forecasting results reaching 10⁻² to 10⁻⁶.
• Showcased optimized time and space complexity compared to existing methods.

Conclusions:

• The Informer model effectively mitigates error accumulation in motor bearing vibration forecasting.
• It offers a promising solution for accurate and efficient prediction, contributing to predictive maintenance and energy saving.