Abstract
Bearings are critical components in mechanical systems; their failure often leads to significant operational disruptions. Rotor unbalance is the most prevalent cause of bearing failure, making accurate state assessment essential for predictive maintenance. This paper proposes a method combining Welch Power Spectral Density (Welch-PSD) estimation and Stacked Auto-Encoders (SAE) to assess the degree of bearing rotor static unbalance. By processing two-way vibration signals from vertical and horizontal acceleration sensors, the Welch-PSD method extracts the vibration power at the working frequency as a primary feature. These features are then fed into an SAE neural network to quantify the unbalance state. Experimental results demonstrate that Welch-PSD accurately isolates fault features, while the SAE effectively maps these features to a precise health index, enabling sensitive detection of varying unbalance degrees.
Introduction
In rotating machinery, bearings are vulnerable components whose health directly impacts system reliability. Rotor unbalance—a condition where the center of mass deviates from the axis of rotation—is the leading cause of excessive vibration and eventual bearing breakdown. While identifying the presence of a fault is common, quantifying the severity of unbalance remains a challenge due to signal noise and non-linear relationships in vibration data. This research addresses this gap by providing a data-driven state assessment framework that translates raw vibration signals into a continuous health index.
Methodology
The proposed diagnostic pipeline consists of two primary stages: feature extraction and state assessment.
Feature Extraction via Welch-PSD
Traditional Fast Fourier Transform (FFT) methods can be sensitive to noise and spectral leakage. To mitigate this, the researchers employed the Welch-PSD method. This approach partitions the vibration signal into overlapping segments, applies windowing, and averages the results to produce a cleaner power spectrum. The core feature extracted is the vibration power at the rotor's working frequency, which serves as the most significant indicator of static unbalance.
State Assessment via Stacked Auto-Encoder (SAE)
The Stacked Auto-Encoder (SAE) is utilized to map extracted features to a health state. The process involves an unsupervised phase where the network learns efficient data representations, followed by a supervised fine-tuning phase. In this latter stage, the model is trained to output a "Health Index" (HI), where a value of 1.0 represents a normal state and 0 represents a critical failure state. This regression-style approach allows for the quantification of degradation rather than simple binary classification.
Experimental Setup
The methodology was validated using a Machinery Fault Simulator (MFS) test-bed. The setup involved an experimental bearing operating at a working frequency of 30 Hz. Vibration signals were collected at a high sampling rate of 40,000 Hz across nine distinct states: one normal state and eight levels of injected rotor unbalance ranging from 94.7 g·mm to 528.7 g·mm.
Results and Discussion
The experimental results confirmed the robustness of the Welch-PSD and SAE framework.
Feature Clarity
A key finding was the high degree of separability achieved by the Welch-PSD method. By fusing data from both vertical and horizontal sensors, the model gained a multi-axis view of the vibration. As shown in the feature clustering analysis, the vibration power at the working frequency formed distinct, non-overlapping groups for each of the nine unbalance levels. This clarity is essential for the subsequent neural network to perform accurate mapping.
Health Mapping Accuracy
The SAE model successfully produced a health index curve that tracked the injected fault levels with high precision. As the physical unbalance increased, the model’s output dropped predictably from 1.0 toward 0. This sensitivity allows the system to distinguish between very fine increments of unbalance, which is critical for early-stage fault detection and informed maintenance scheduling.
Conclusion
This paper demonstrates that the combination of Welch-PSD and SAE provides a powerful tool for the state assessment of bearing rotor unbalance. The method effectively simplifies complex vibration data into actionable health metrics. Future work will focus on testing this framework in more chaotic industrial environments and expanding the model to handle compound failures where multiple types of bearing defects occur simultaneously.
Summary updated