.png){kind=icon}
.png){kind=icon}
Return
A Novel and Robust Lightweight Deep Learning Framework for Bearing Fault Detection in Diverse Environments: Comprehensive Evaluation, Comparison Study, and Experimental Verification
.png){kind=icon}
PRE
AI
G
T
N
DOI：10.1109/JESTIE.2025.3647729.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diagnosing bearing faults, which are the most critical and frequent type of fault that can lead to other faults in rotating electrical machines, has been a challenge for engineers and researchers in recent years. Several deep learning (DL) methods have been proposed to detect bearing faults with high accuracy. However, the practical implementation of these models must be considered. Two significant challenges in the real-world execution of these methods are: 1) Theoretical methods often have many parameters, making them heavy-weighted and slow, and 2) environmental noise is another significant issue that affects accuracy of detection. In DL, vision transformers (ViT) have gained popularity among researchers in computer vision tasks. Unlike convolutional neural networks (CNNs), which learn local features, ViTs can learn global representations. However, ViTs are heavier than CNNs. To tackle the challenges in the practical implementation of bearing fault detection and leverage the advantages of both ViT and CNN, the MobileViT, a lightweight and mobile-friendly ViT, is utilized for bearing fault detection for the first time in literature. This article evaluates various types of bearing faults and fault severities in multilevel loads. To achieve this, 1-D vibration signals from CWRU dataset were converted to 2-D time-frequency image samples using continuous wavelet transformation. MobileViT was trained on time-frequency images to learn representative features and classify them. To consider the noise condition, the model's performance in the range of -4 to 4 signal-to-noise ratio (SNR) was investigated. The proposed approach was also validated on real bearing damages from the Paderborn dataset to assess generalization capability. The results are compared with other methods and demonstrate the effectiveness and applicability of the proposed method.
Keywords:
Vibrations
Feature extraction
Convolutional neural networks
Multiresolution analysis
Fault detection
Transformers
Training
Time-frequency analysis
Noise measurement
Computer vision
Bearing fault diagnosis
convolutional neural networks (CNNs)
deep learning (DL)
vision transformers (ViT)
Journal
I
IF:
0
Papers:
138
Citations:
0
Organization
Cited Papers