TL;DR
This paper introduces an enhanced signal decomposition technique (ECEEMDAN) for more accurate fault diagnosis in rolling bearings, especially under variable speed conditions. By intelligently selecting decomposition parameters and combining statistical features into a comprehensive factor, the method effectively identifies subtle fault signatures that traditional approaches miss.
What Problem Does This Paper Address?
Rolling bearings are critical components in rotating machinery, and their failure can lead to significant downtime and safety hazards. Detecting faults early is crucial for predictive maintenance. However, vibration signals from bearings are often non-linear and non-stationary, making them difficult to analyze. Traditional time-domain, frequency-domain, and even existing time-frequency methods struggle to accurately extract fault features, especially in the presence of noise and under varying operating conditions like speed changes. Existing decomposition methods like EMD, EEMD, and CEEMDAN suffer from issues like mode mixing, residual noise, and spurious modes, or require computationally expensive parameter tuning.
Methods
The core of this paper is the Enhanced Complementary Empirical Mode Decomposition with Adaptive Noise (ECEEMDAN). This method builds upon CEEMDAN but introduces two key improvements:
- Adaptive Parameter Selection: Instead of relying on empirical rules, ECEEMDAN determines the optimal amplitude of added white noise (AAWN) by analyzing the mutual information (MI) between adjacent intrinsic mode functions (IMFs) at different decomposition levels. This ensures that the decomposition effectively separates signal components while minimizing spurious modes.
- Fixed Ensemble Trials: The number of ensemble trials (ET) is fixed at two. This significantly reduces computational cost compared to traditional methods that use hundreds of trials, without compromising decomposition quality.
Once the signal is decomposed into IMFs, the paper proposes a Comprehensive Factor (CF) to identify the most informative IMF, termed the "intrinsic information mode" (IIM). The CF combines several statistical time-domain features (shape factor, crest factor, impulse factor, kurtosis, skewness, and latitude factor) by giving them equal weight. This approach leverages the strengths of each individual feature while mitigating their weaknesses, leading to a more robust identification of the IIM. Finally, the envelope spectrum of the identified IIM is analyzed to extract fault features.
Datasets
The paper utilizes two main sources of data for evaluation:
- Simulated Signals: Various simulated signals are used, including:
- A free-noise signal with three distinct frequency components.
- Noisy signals ('blocks', 'bumps', 'heavy sine waves', 'quadchirps') with varying Signal-to-Noise Ratios (SNRs).
- A simulated vibration signal (SVS) designed to mimic bearing faults, with added white Gaussian noise.
- Experimental Data:
- Case Western Reserve University (CWRU) Bearing Data: This dataset includes vibration signals from bearings with inner and outer raceway faults under controlled conditions.
- Real Measurement Data: Vibration signals were acquired from a custom-built experimental rig featuring a motor with a defective bearing and a healthy bearing, tested at different rotation speeds.
Key Insights and Ideas
- Mutual Information for Adaptive Decomposition: Using MI to guide the selection of AAWN is a novel approach that directly addresses the challenge of mode mixing and spurious modes in EMD-based methods. It provides a data-driven way to find the optimal decomposition level.
- Comprehensive Factor for IIM Selection: Combining multiple statistical features into a single CF is more effective than relying on any single feature. This is crucial because different statistical features are sensitive to different types of faults or fault severities.
- ECEEMDAN's Efficiency: Fixing ET to two significantly speeds up the decomposition process, making the method more practical for real-time applications.
- Superiority in Variable Speed Conditions: The method demonstrates remarkable effectiveness in identifying fault frequencies even when the bearing operates at varying speeds, a common challenge in industrial settings.
Key Results
- Improved Decomposition: ECEEMDAN consistently produced fewer IMFs compared to traditional EMD variants (EEMD, CEEMD, CEEMDAN) for the same signals, indicating better suppression of spurious modes.
- Enhanced Fault Feature Extraction: The proposed ECEEMDAN-CF method successfully identified fault frequencies and their harmonics, including those that were difficult to detect with other methods, especially in noisy conditions or under variable speeds. For instance, in the SVS signal, the IIM identified the third harmonic (3fc) which was challenging for the original signal.
- Computational Efficiency: ECEEMDAN showed a significant reduction in computation time compared to CEEMDAN (e.g., 0.48s vs. 9.16s for the inner raceway fault signal).
- Robustness to Noise: The method maintained strong decomposition capabilities and accurate fault identification across a range of SNRs in noisy signal tests.
- Real-World Validation: Experiments on both CWRU data and the custom rig confirmed the method's ability to accurately detect bearing faults under realistic operating conditions.
Limitations and Future Work
- Decomposition Level Investigation: The paper acknowledges that the optimal decomposition level (Lopt) needs to be investigated for different applications. While the MI-based approach helps determine it, further research could refine this process.
- Generalizability: While the method shows promise across various simulated and experimental scenarios, its performance on a wider range of bearing types and fault severities could be explored.
- Real-time Implementation: Further work could focus on optimizing the algorithm for even faster real-time implementation in industrial monitoring systems.
Summary updated