Abstract
Rolling bearings are critical components in industrial machinery, yet their automated fault diagnosis is often hampered by heavy background noise and varying load conditions. This paper introduces the Adaptive Anti-Noise Neural Network (AAnNet), an end-to-end framework designed to identify bearing failures directly from raw vibration signals. By integrating convolutional layers with Exponential Linear Units (ELU), Gated Recurrent Units (GRU), and a temporal attention mechanism, AAnNet achieves state-of-the-art accuracy without requiring manual feature engineering. Experimental results on benchmark and custom datasets demonstrate superior robustness at low Signal-to-Noise Ratios (SNR) and improved stability under shifting mechanical loads.
Introduction
In modern manufacturing, the failure of rolling bearings can lead to catastrophic machine breakdowns. While deep learning has advanced automated diagnosis, models often struggle when transitioning from controlled laboratory environments to the "messy" reality of the factory floor. Two primary challenges persist: heavy background noise that masks fault signatures and varying load conditions that alter a machine's vibration profile. Traditional solutions rely on manual denoising and handcrafted statistical features. AAnNet aims to eliminate this manual intervention through an adaptive, robust architecture.
Methodology
The AAnNet architecture employs a multi-stage approach to process raw vibration data:
1. Robust Feature Extraction
The input layer utilizes a Random Sampling Strategy to simulate data uncertainty, forcing the network to learn generalized features. The encoder stage uses Convolutional Neural Networks (CNN) with wide kernels to capture a broad receptive field. Standard ReLU activations are replaced with Exponential Linear Units (ELU), which improve noise robustness by allowing negative values and reducing bias shifts.
2. Temporal Modeling with GRU
Because vibration is a time-series signal, the model incorporates Gated Recurrent Units (GRUs) to capture temporal dependencies. The GRU effectively "listens" to the evolution of the signal over a full motor rotation, identifying patterns that static filters might miss.
3. Attention Mechanism
To further refine the output, an Attention Mechanism is applied to the GRU hidden states. This layer automatically weights the most relevant time steps, allowing the model to focus on specific "clicks" or "rumbles" indicative of a fault while ignoring irrelevant noise.
Experimental Validation
The model was validated using two datasets:
- CWRU Dataset: The industry-standard benchmark from Case Western Reserve University, covering 10 classes of bearing health.
- QPZZ-II Dataset: A custom dataset generated to test the model's universality across different hardware configurations.
Key Results and Discussion
- Noise Robustness: At a highly noisy -10 dB SNR, AAnNet maintained high accuracy, significantly outperforming established baselines like TICNN and WDCNN. The use of ELU and random sampling proved critical for stability in these conditions.
- Load Adaptation: In cross-domain tests (training on one load and testing on another), the AAnNet backbone achieved an average accuracy of 61.68%, nearly doubling the performance of traditional Support Vector Machines (36.61%).
- Architectural Insights: Visualization of the first-layer kernels revealed that the CNN effectively learns to perform a Power Spectrum analysis similar to a Fast Fourier Transform (FFT). Furthermore, t-SNE mapping showed that while the CNN aggregates features, the GRU layers act as the primary decoders that separate different fault types into distinct clusters.
Conclusion
AAnNet provides a robust, end-to-end solution for industrial fault diagnosis. By treating noise as a condition to be mastered rather than a nuisance to be removed, the framework offers high reliability under fluctuating real-world conditions. Future work will focus on improving the flexibility of the attention mechanism for extreme load shifts and addressing challenges posed by imbalanced datasets.
Summary updated