Fault Feature Extraction for Reciprocating Compressors Based on Underdetermined Blind Source Separation.

Area of Science:

• Engineering
• Signal Processing
• Data Science

Background:

• Vibration signals in engineering often contain outliers, degrading source signal separation accuracy.
• Accurate mixing matrix estimation is vital for reliable underdetermined blind source separation (UBSS).
• The K-means algorithm is sensitive to initial cluster centers and outliers, impacting separation performance.

Purpose of the Study:

• To propose a novel two-stage clustering method combining hierarchical clustering and K-means.
• To enhance the reliability of mixing matrix estimation in UBSS.
• To improve the accuracy of source signal recovery from noisy vibration data.

Main Methods:

• A two-stage clustering approach integrating hierarchical clustering and K-means is developed.
• Hierarchical clustering identifies initial cluster centers, followed by outlier removal using cosine distance.
• Improved K-means estimates the mixing matrix, and source recovery is performed using the least squares method.

Main Results:

• The proposed method effectively addresses K-means sensitivity to initial centers and outliers.
• Outlier elimination improves the reliability of the estimated mixing matrix.
• Both simulations and reciprocating compressor fault experiments validate the method's effectiveness.

Conclusions:

• The developed two-stage clustering method offers a robust solution for UBSS in the presence of signal outliers.
• This approach significantly enhances mixing matrix estimation accuracy.
• The method demonstrates practical applicability in engineering, particularly for fault diagnosis using vibration signals.