TL;DR
This paper introduces the Band Relevance Factor (BRF), a novel method for automatically identifying crucial frequency bands in vibration signals from rotating machinery. By analyzing spectral entropy and RMS values, BRF provides a ranked list of relevant bands, simplifying fault diagnosis and feature extraction for AI models.
What Problem Does This Paper Address?
Rotating machinery is critical in many industries, and monitoring its health through vibration analysis is essential for preventing costly failures. However, raw vibration signals are often noisy and complex, making it challenging to extract meaningful information. Existing methods for identifying "informative frequency bands" (IFB) often focus solely on impulsive excitations or require manual feature extraction and selection, lacking automation and efficiency. This leaves a gap in automatically pinpointing all relevant frequency bands that reflect the machine's dynamic behavior and potential faults, beyond just impulsive events.
Methods
The core of the proposed method is the Band Relevance Factor (BRF). It operates in several steps:
- Raw Signal Entropy Calculation: The initial step involves calculating the spectral entropy of the raw vibration signal. A threshold of -3 dB is used: signals with entropy above this are considered noise and irrelevant.
- Signal Filtering and RMS Calculation: Relevant signals are then filtered into multiple frequency bands (k-levels). For each band and the original signal, the Root Mean Square (RMS) value is calculated. The difference between the RMS of the original and filtered band is termed the Correction Factor (rmsdiff). This factor accounts for the energy present in each band.
- Entropy Difference Factor (Sdiff): The difference in spectral entropy between the original signal and each filtered band is calculated, adjusted by a constant to shift the decision threshold. This measures the change in signal order/uncertainty.
- Band Relevance Factor (BRF) Calculation: The BRF is computed by dividing the Entropy Difference Factor (Sdiff) by the Correction Factor (rmsdiff). Positive BRF values indicate a relevant band, while negative values suggest irrelevance. A higher BRF signifies greater relevance.
- Relevance Ranking and Heatmap: The BRF values are used to create a relevance ranking for each frequency band. A heatmap visualization is also provided to offer a visual representation of these rankings across different frequency bands and levels of signal decomposition.
Datasets
The methodology was validated using three datasets:
- Case 1: Synthetic Dataset: A synthetic signal was generated to mimic real-world rotating machinery conditions, including various harmonic frequencies and controlled levels of Gaussian noise (0 dB, 6 dB, 12 dB, 24 dB SNR). This dataset allowed for precise control over signal characteristics.
- Case 2: Bearing Dataset: A publicly available dataset of bearing vibration signals was used, including normal operating conditions and outer race faults.
- Case 3: Mechanical Faults Dataset: A dataset developed by the authors, featuring a test bench with introduced faults such as unbalance, misalignment, and mechanical looseness, under normal operating conditions.
Key Insights and Ideas
- Beyond Impulsive Excitation: Unlike many IFB methods that focus on impulsive events, BRF aims to identify all relevant frequency bands, including those related to the machine's fundamental rotation frequency and other harmonic components that indicate dynamic behavior or specific faults.
- Entropy as a Measure of Order/Irregularity: The paper leverages spectral entropy, where higher entropy generally indicates more noise or irregularity, and lower entropy suggests more ordered (harmonic) components. This is crucial for distinguishing between healthy and faulty states.
- RMS for Energy Context: The inclusion of RMS as a "Correction Factor" is vital. It ensures that bands with high energy (even if not particularly "ordered" in an entropic sense) are appropriately weighted, preventing the method from being misled by noise with significant amplitude.
- Automated Relevance Ranking: The BRF provides a quantitative ranking, enabling automated feature selection for machine learning models and offering clear guidance for manual analysis.
- Robustness to Noise: The BRF demonstrates a degree of robustness to noise, particularly in distinguishing relevant bands when compared to RMS alone, especially in higher noise conditions.
Key Results
- The BRF method successfully identified the synthetically inserted harmonic frequencies across various noise levels in Case 1.
- In Case 2 (bearing fault), BRF identified relevant bands in both normal and fault conditions, including frequencies below 2000 Hz that might be missed by methods focusing solely on high-frequency excitations.
- For Case 3 (mechanical faults), BRF accurately highlighted the rotation frequency band as most relevant for the unbalance condition.
- Comparisons with RMS alone showed that BRF provided a more discerning selection of relevant bands, particularly in distinguishing between signal and noise, and in identifying subtle fault progression indicators.
- In mixed noise conditions (0 dB SNR), BRF yielded a null ranking, correctly indicating that relevant bands could not be reliably separated from noise, a crucial insight for avoiding irrelevant feature extraction.
Limitations and Future Work
- Band Size Sensitivity: The paper notes that when bands become very small, BRF can sometimes classify noise as relevant due to the limited frequency content within that small band. This is mitigated by analyzing rankings and heatmaps.
- Future Work: The authors propose exploring the BRF value for anomaly detection and drift concept detection over time. They also suggest investigating its use as a monitoring feature for component end-of-life prediction and studying its behavior in time series from different applications (e.g., EEG).
Summary updated