TL;DR
This paper introduces a novel tacholess order tracking method that combines Inverse Short-Time Fourier Transform (ISTFT) and Singular Value Decomposition (SVD) for improved bearing fault diagnosis in machines with varying rotational speeds. By effectively filtering and refining the signal components related to shaft rotation, the method accurately estimates the shaft's phase, enabling precise resampling for robust fault detection.
What Problem Does This Paper Address?
Diagnosing faults in rotating machinery is crucial for safe and economical operation. A significant challenge arises when the machine's rotational speed varies, causing spectral line smearing in traditional analysis methods. Order tracking is a well-established technique to mitigate this, but it typically requires a tachometer signal to measure the shaft's speed. When a tachometer signal is unavailable (due to cost or space constraints), Tacholess Order Tracking (TOT) methods become essential. Existing TOT methods often focus on accurately estimating the instantaneous frequency. However, this paper addresses the need for a robust TOT method that can effectively separate and refine the signal component related to shaft rotation, even in the presence of noise, to enable accurate phase extraction for fault diagnosis.
Methods
The proposed method integrates several key techniques:
- Short-Time Fourier Transform (STFT): Used to obtain the time-frequency spectrum of the raw vibration signal, allowing for visualization and initial analysis of frequency components over time.
- Component Selection and Filtering: A crucial step where the strongest harmonic related to the shaft rotation frequency (or gear meshing frequency) is identified and isolated in the time-frequency domain. This filtering is approximate, aiming to isolate the target component without requiring precise instantaneous frequency estimation.
- Inverse Short-Time Fourier Transform (ISTFT): Applied to the filtered time-frequency spectrum to reconstruct the corresponding time-domain signal of the selected component.
- Singular Value Decomposition (SVD): Used to refine or denoise the time-domain signal obtained from ISTFT. SVD decomposes the signal into sub-signals, and by selecting the most significant ones, it effectively filters out noise and improves the accuracy of the component.
- Phase Extraction: The instantaneous phase of the refined signal is extracted.
- Resampling: The extracted phase information is used to resample the original signal in the angular domain, effectively removing the smearing caused by speed variations. This is achieved using spline interpolation.
The novelty lies in the combination of ISTFT for initial filtering and SVD for subsequent refinement, making the time-frequency domain filtering step less sensitive to precise frequency estimation.
Datasets
The method was validated using real vibration signals collected from a large-scale test rig of a high-speed train traction system. This test rig included components like motors, gearboxes, and brake motors, and featured bearings with known fault frequencies. Two case studies were presented:
- Case Study 1: Involved an artificial spall on the outer ring of a FAG-804989 bearing, with rotational speeds varying from 150 to 970 rpm.
- Case Study 2: Involved an artificial spall on the inner ring of an SKF NU215 bearing, with rotational speeds varying from 500 to 4950 rpm.
The paper does not introduce new benchmark datasets but utilizes real-world experimental data from a relevant industrial application.
Key Insights and Ideas
- SVD as a Refinement Tool: The paper highlights SVD's effectiveness not just for decomposition but as a powerful denoising and refinement step after initial time-frequency filtering, significantly improving phase extraction accuracy.
- Relaxed Time-Frequency Accuracy Requirement: By using SVD for refinement, the initial time-frequency filtering step can be less precise, simplifying the process and making it more robust to variations.
- Importance of Component Selection: The accurate selection of the target harmonic related to shaft rotation is critical for the success of the entire method.
- Noise Reduction through SVD: The study demonstrates that increasing the number of rows in the Hankel matrix for SVD decomposition can effectively reduce the influence of noise on phase extraction, especially at lower SNRs.
- Experimental Validation: The method's efficacy is proven on real-world vibration data from a high-speed train traction system, showcasing its practical applicability for bearing fault diagnosis.
Key Results
- The proposed tacholess order tracking method successfully extracts the shaft phase with high accuracy, closely matching the actual phase measured by a tachometer in experimental cases.
- The method effectively alleviates spectral line smearing caused by varying rotational speeds, enabling clearer identification of bearing faults.
- In Case Study 1 (outer ring defect), the Square Envelope Spectrum (SES) of the resampled signal clearly identified the fault by showing peaks at the first two harmonics of the BPFO.
- In Case Study 2 (inner ring defect), the SES of the resampled signal also successfully identified the defect with a peak at BPFI.
- The study showed that for a maximum phase error of less than 5% in one cycle, the relative frequency ratio should generally be kept below 3.0, and this limit is influenced by the initial frequency.
- Noise analysis indicated that with an SNR of approximately 0 dB, the maximum phase error in a cycle can be kept below 5% when using an appropriate number of rows for SVD.
Limitations and Future Work
- Computational Memory Requirements: The paper acknowledges that the SVD operation can be computationally intensive, requiring significant memory, which is a drawback for real-time applications or systems with limited resources.
- Future Work: The authors plan to apply this method to monitor the condition of bearings in train traction systems, suggesting a focus on practical implementation and continuous monitoring. Further research could explore optimizations for the SVD step to reduce computational load and enable real-time deployment.
Summary updated