TL;DR
This paper introduces a novel end-to-end fault diagnosis scheme for rolling bearings that leverages capsule networks (CNs) and fused two-directional vibration signals. This approach effectively addresses the challenge of varying rotational speeds, a common issue in real-world machinery, by automatically extracting invariant features and achieving high diagnostic accuracy.
What Problem Does This Paper Address?
Rolling bearings are critical components in rotating machinery, and their failure can lead to catastrophic consequences. Effective fault diagnosis is therefore essential for ensuring safety and reducing costs. While vibration analysis is a widely used method, traditional intelligent diagnosis techniques often rely on manual feature extraction, which can be time-consuming and require expert knowledge. Deep learning methods have shown promise, but many existing approaches assume that training and testing data share similar distributions. This assumption is often violated in practice, especially for rotating machinery where rotational speeds can vary significantly. These speed variations introduce non-stationarity into vibration signals, making it difficult to accurately diagnose faults. Existing methods to mitigate this often require manual speed information or complex transfer learning setups. This paper aims to automatically eliminate the influence of different rotational speeds without explicit speed information.
Methods
The core of this paper is an end-to-end fault diagnosis scheme that combines two key innovations:
- Capsule Networks (CNs): Unlike traditional Convolutional Neural Networks (CNNs) that can lose spatial information through pooling, CNs are designed to preserve and leverage the spatial relationships between features. This is achieved by representing features as vectors, where the vector's length signifies the probability of a feature's existence and its orientation encodes its properties. The paper utilizes a CN architecture with a convolutional layer, primary capsule layer, digit capsule layer, and a decoder for reconstruction. The "dynamic routing" algorithm is central to the CN, allowing capsules to learn to route information to higher-level capsules based on agreement, rather than relying solely on backpropagation.
- Two-Directional Signal Fusion: To further combat the influence of varying speeds, the authors fuse horizontal and vertical vibration signals from two orthogonal sensors. This fusion is performed using two methods: "channel overlay" (concatenating signals along the channel dimension) and "stack banding" (concatenating along the width dimension). These fused signals are then treated as a two-dimensional input for the CN.
The overall process involves collecting vibration signals, preprocessing them (normalization and segmentation), fusing the two-directional signals, and then inputting the fused data into the trained CN for fault diagnosis.
Datasets
The experiments were conducted using data collected from an MFS-MG test rig at the University of Electronic Science and Technology of China (UESTC). The dataset includes vibration signals from rolling bearings under four health states: normal, inner race fault, rolling element fault, and outer race fault. Data was collected at three different rotational speeds: 800 rpm, 1500 rpm, and 1800 rpm. This creates distinct datasets for each speed, allowing for evaluation under varying conditions. The paper also compares its proposed method with a Wide Deep Convolutional Neural Network (WDCNN) using the same datasets.
Key Insights and Ideas
- CNs Excel at Capturing Spatial Relationships: The paper highlights the inherent advantage of CNs in preserving the spatial relationships between features, which is crucial for understanding complex vibration patterns in machinery. This ability is key to extracting invariant features that are less sensitive to rotational speed changes.
- Two-Directional Fusion Enhances Robustness: Fusing horizontal and vertical vibration signals significantly improves the diagnostic performance, especially when dealing with large speed gaps between training and testing data. This suggests that combining information from different perspectives provides a more comprehensive and robust representation of the bearing's health.
- Automatic Speed Invariance: The proposed CN-based scheme, by leveraging fused two-directional signals, demonstrates an ability to automatically learn features that are invariant to rotational speed variations, eliminating the need for explicit speed measurements or complex domain adaptation techniques.
- Data Fusion Methods Have Minor Impact: Interestingly, the study found that the specific method of fusing the two-directional signals (channel overlay vs. stack banding) had a relatively minor impact on the overall fault diagnosis results, suggesting that the core benefit comes from using two-directional data with the CN.
Key Results
- High Accuracy under Same Speed: When training and testing data were from the same rotational speed, the CN model achieved near-perfect accuracy (100 ± 0.00%) across all three speeds, outperforming the WDCNN.
- Significant Improvement under Different Speeds: When tested under different rotational speeds (e.g., training at 800 rpm, testing at 1500 rpm or 1800 rpm), the CN model with two-directional signal fusion achieved significantly higher accuracies compared to the WDCNN. For instance, in the "a→c" scenario (800 rpm to 1800 rpm), the CN achieved an average accuracy of 75.53% ± 3.88%, while the WDCNN achieved only 46.41% ± 7.24%.
- Enhanced Performance with Two Sensors: Using two-directional sensors (fusion methods B and C) consistently led to higher accuracies compared to using a single sensor (method A), especially under varying speeds.
- Training Time Trade-off: While the CN model with two-directional fusion sometimes had a slightly longer training time than the WDCNN, the significant improvement in diagnostic accuracy, particularly under challenging variable speed conditions, was deemed a worthwhile trade-off.
Limitations and Future Work
- Fluctuating Speeds: The paper acknowledges that the current study focused on distinct rotational speeds. For future work, the authors propose investigating the performance of their proposed scheme under more realistic scenarios of fluctuating speeds.
- Data Fusion Impact: While the impact of different fusion methods was found to be minor, further exploration into optimal fusion strategies could potentially yield even better results.
Summary updated