Multidimensional fast nonlinear blind deconvolution network for bearing compound features extraction.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Condition Monitoring

Background:

• Single bearing faults can precipitate new faults due to uneven stress and abnormal loads.
• Weak fault features in compound faults are often masked by existing fault signatures or interference.
• Accurate identification of compound faults is crucial for preventing catastrophic failures in machinery.

Purpose of the Study:

• To develop a novel method for the separation and extraction of compound fault features.
• To address the challenge of weak and aliased fault signatures in complex fault scenarios.
• To enhance the robustness and accuracy of fault diagnosis in rotating machinery.

Main Methods:

• Extension of fast nonlinear blind deconvolution (FNBD) to multidimensional FNBD (MFNBD) for composite feature decoupling.
• Introduction of uniform multidimensional initialization to guide multi-feature extraction convergence.
• Development of trimmed envelope spectrum kurtosis for eliminating irrelevant components based on harmonic distribution.
• Incorporation of adaptive nonlinear transformation and filter waveform penalty into the MFNBD-net framework.

Main Results:

• MFNBD-net demonstrated superior performance in multidimensional feature decoupling compared to existing methods.
• The proposed method effectively separated and extracted weak fault features from complex interference.
• Simulations and experimental results validated the robustness and accuracy of MFNBD-net.

Conclusions:

• MFNBD-net is a promising tool for extracting composite fault features in condition monitoring.
• The method offers significant advantages in handling complex fault scenarios with multiple simultaneous issues.
• This advancement contributes to more reliable and effective machinery fault diagnosis.