Boosted Convolutional Neural Network Algorithm for the Classification of the Bearing Fault form 1-D Raw Sensor Data

Area of Science:

• Engineering
• Data Science
• Renewable Energy

Background:

• Wind farms are expanding, leading to increased maintenance challenges.
• Bearing failures account for approximately 40% of all wind turbine failures.
• There is a critical need for advanced, non-intrusive predictive maintenance strategies.

Purpose of the Study:

• To introduce a novel predictive maintenance method for wind turbine bearings.
• To enhance failure detection accuracy and reduce computational load.
• To demonstrate a real-time, less intrusive diagnostic approach.

Main Methods:

• A modified neural direct classifier utilizing raw accelerometer measurements.
• Directly processing tabular data without feature extraction from spectrograms.
• Real-time analysis without converting signals to time-frequency spectrograms.

Main Results:

• Achieved 99.32% precision on the validation set and 96.3% during bench testing.
• Outperformed traditional time-frequency spectrogram classification methods (97.76% validation, 90.8% real-world).
• Demonstrated superior damage prediction compared to convolutional networks in vibration spectrum analysis.

Conclusions:

• The proposed method significantly improves wind turbine bearing failure detection accuracy.
• Directly analyzing raw data reduces computational costs and enhances efficiency.
• This approach represents a significant advancement in predictive maintenance for renewable energy infrastructure.