Abstract
Due to the disadvantages that rely on prior knowledge and expert experience in traditional order analysis methods and deep learning cannot accurately extract the features in time-varying conditions. A fault diagnosis method for rotating machinery under time-varying conditions based on tacholess order tracking (TOT) and deep learning is proposed in this paper. Firstly, frequency domain periodic signals and estimated speed information are obtained by order tracking. Secondly, the frequency domain periodic signal is speed normalized using the estimated speed information. Finally, normalized features are extracted by deep learning network to form feature vector. The feature vector is fed into a softmax layer to complete fault diagnosis of the gearbox. The fault diagnosis of the gearbox results are compared with other traditional methods and show that the proposed fault diagnosis method can effectively identify the faults and obtain higher fault diagnosis accuracy under time-varying speed.
Introduction
In industrial environments, machinery such as gearboxes and turbines frequently operate under time-varying speeds. Traditional diagnostic tools like the Fast Fourier Transform (FFT) are often inadequate in these scenarios because frequency components "smear" across the spectrum as rotational speed fluctuates. While deep learning models have shown potential, they struggle to generalize when input data distributions shift due to these speed changes. Furthermore, many industrial setups cannot accommodate physical tachometers for speed measurement. This paper addresses these issues by proposing a tacholess order tracking (TOT) framework combined with a Long Short-Term Memory (LSTM) network.
Methodology
The proposed diagnostic framework bridges signal processing physics with deep learning through a multi-stage pipeline.
Tacholess Order Tracking (TOT) and Normalization
The first stage involves estimating the "virtual speed" directly from vibration signals using a Gabor-transform-based algorithm. This is followed by equal-angle resampling, which converts the signal from the time domain to the angular domain, effectively "freezing" non-stationary signals into stationary periodic ones.
To further stabilize the input for the neural network, Rotational Speed Normalization (RSN) is applied. Since vibration amplitude naturally increases with speed due to centrifugal force ($f = mw^2r$), the signal amplitude is divided by the square of the instantaneous speed.
LSTMBN Model Architecture
The processed data is fed into an LSTM network enhanced with Batch Normalization (BN). The LSTM architecture is specifically designed to capture temporal dependencies in sequence data, while BN layers stabilize the training process against internal covariate shifts.
Experimental Results
The approach was validated using simulated signals and real-world data from a Siebenhaar CLP220 three-stage gearbox used in portal cranes. The dataset comprised 101 groups of samples covering healthy states and common failures like tooth surface and bearing wear.
The proposed TOT-LSTMBN model achieved a diagnostic accuracy of 99.78%, significantly outperforming several benchmarks:
- TOT-CNN: 96.74%
- TOT-SVM: 89.71%
- Standard LSTMBN (without Order Tracking): 83.70%
These results highlight that while LSTMs are effective for time-series tasks, the Order Tracking (TOT) preprocessing step is the critical factor for reliability under variable speeds.
Conclusion
This research demonstrates that combining physics-informed preprocessing with deep learning creates a robust, tacholess solution for machinery fault diagnosis. Future work will focus on applying this method to more complex rotor-bearing systems and diagnosing distributed faults where multiple components fail simultaneously.
Summary updated