TL;DR
This paper introduces a novel method called Frequency Exchange and Re-Scaling Stochastic Resonance (FERSR) to overcome limitations in detecting weak, multi-frequency fault signals in mechanical systems. FERSR effectively handles signals with low sampling ratios, a common challenge for existing Stochastic Resonance (SR) techniques, by exchanging frequency information and then applying SR.
What Problem Does This Paper Address?
Diagnosing faults in rotating machinery often relies on identifying characteristic fault frequencies and their harmonics. However, these signals can be very weak and buried in significant background noise. Traditional methods like linear amplification fail because they boost both the signal and the noise, leaving the Signal-to-Noise Ratio (SNR) unchanged. Stochastic Resonance (SR) offers a way to enhance weak signals in noise, but classical SR methods have strict "small-parameter conditions" that are often violated by real-world fault signals, especially those with high frequencies or when sampling rates are not sufficiently high relative to the signal frequency (low sampling ratio). Existing SR variants that attempt to address this, like Frequency Re-scaling Stochastic Resonance (FRSR), still struggle with low sampling ratios and can lead to divergent results. This paper aims to provide a more robust and practical solution for detecting multi-frequency weak fault signals, particularly when dealing with low sampling ratios.
Methods
The core of the proposed method, FERSR, involves two main stages:
- Frequency Exchange: This is achieved using a combination of filtering techniques and Single Sideband (SSB) modulation.
- Filtering: The original signal is first processed by a filter bank. This includes band-pass filters to extract local information around each target characteristic frequency and a low-pass filter for the reference low frequency. A high-pass filter and band-stop filters are used to obtain the residual signal.
- SSB Modulation: The extracted local characteristic frequency information is then "exchanged" with the reference low-frequency information. This is done by modulating the target signals with a carrier frequency derived from the difference between the target frequency and the reference frequency. This process effectively shifts the high-frequency target signals to a lower frequency band, making them amenable to SR processing. The residual signal is then added back to reconstruct the frequency-exchanged signal.
- Re-Scaling Stochastic Resonance (SR): After frequency exchange, the signals are rescaled to meet the small-parameter conditions required by classical SR. The rescaled signals are then fed into an SR system. The output of the SR system is then processed through frequency recovery to obtain the original fault frequencies.
The paper highlights that this approach overcomes the "sampling ratio" limitation of FRSR and ensures that multi-frequency target signals can be processed to meet the small-parameter conditions.
Datasets
The paper does not introduce new benchmark datasets. Instead, it uses:
- Simulated Data: A synthetic multi-frequency signal composed of three sinusoidal signals with arbitrary frequencies (42 Hz, 95 Hz, 243 Hz) and Gaussian white noise was generated for initial testing.
- Experimental Data:
- Rolling Bearing Outer Ring Fault: Data was collected from a high-speed train rolling bearing test rig with a simulated outer ring fault. The characteristic frequency was calculated to be around 73 Hz.
- Rotor Shaft-Bending Fault: Data was collected from a sliding-bearing test rig simulating shaft bending. The rotating frequency was 28 Hz, and the fault signal components were attenuated due to damping.
Key Insights and Ideas
- Frequency Exchange as a Key Enabler: The core innovation is the "frequency exchange" step. By shifting high-frequency fault components to a lower, manageable frequency band, the method bypasses the need for extremely high sampling rates or the limitations of existing SR methods that struggle with high frequencies.
- Overcoming Low Sampling Ratio Limitations: The FERSR method is specifically designed to handle signals where the sampling frequency is not much larger than the signal frequency (low sampling ratio). This is a critical practical problem in many industrial monitoring scenarios.
- Multi-Frequency Handling: The filter bank approach allows for the simultaneous extraction and processing of multiple fault frequencies and their harmonics, which is crucial for comprehensive fault diagnosis.
- Practical Applicability: The method's validation on real experimental data from bearing and shaft fault diagnosis demonstrates its effectiveness in real-world engineering problems.
Key Results
- Simulation: FERSR successfully detected three simulated multi-frequency signals (42 Hz, 95 Hz, 243 Hz) with low sampling ratios (e.g., 1000/42 ≈ 23.8), which FRSR failed to do without divergence.
- Rolling Bearing Fault: FERSR accurately identified the first three harmonics of the outer ring fault frequency (around 73 Hz) from experimental data, which were previously obscured by noise. The method showed superior performance compared to FRSR and Singular Value Decomposition (SVD) with envelope analysis.
- Shaft-Bending Fault: FERSR detected the first four harmonics of the shaft-bending fault frequency (28 Hz, 56 Hz, 83 Hz, 112 Hz) from weakened experimental signals, demonstrating its ability to extract very weak fault features.
Limitations and Future Work
- Prior Knowledge Dependency: The current FERSR method requires some prior knowledge of the target fault frequencies to design the filter banks. This makes it less convenient than some adaptive methods.
- Future Work: The authors suggest incorporating intelligent algorithms (e.g., adaptive methods) into the FERSR framework to make it more adaptive and automated for weak multi-frequency signal detection and incipient fault diagnosis. This would address the current limitation of requiring prior knowledge.
Summary updated