Efficient Gearbox Fault Diagnosis: A Leap in Precision

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jan 12 2024

In an article recently published in the journal Scientific Reports, researchers presented a gearbox fault diagnosis method based on transfer learning (TL) and a lightweight channel attention mechanism.

Gearbox fault diagnosis challenges

In mechanical equipment, the gearbox is a critical transmission component that operates in variable and complex conditions. Effective and accurate gearbox fault diagnosis can decrease the frequency of major accidents and improve the security and reliability of equipment operation.

Fault pattern recognition and signal feature extraction are the key aspects of the gearbox fault diagnosis methods. In the era of big data, fault diagnosis involves many diverse and complex collected signals. Conventional fault identification methods are ineffective in the big data era, which has led to an increasing focus on intelligent gearbox fault identification. In recent years, deep learning (DL) has increasingly been utilized extensively in fault diagnosis.

In DL methods, convolutional neural network (CNN) has significant two-dimensional (2D) image classification advantages and is commonly used for mechanical equipment fault diagnosis. Although the DL algorithm can extract fault features adaptively and possesses robust fault classification performance, it requires a substantial number of samples for fault diagnosis. However, the monitored gearbox signals are primarily normal operation data and only a few fault samples are available, which adversely affects the generalization capability and recognition precision of neural networks.

The TL method can effectively address such problems as a TL model eliminates the need for numerous labeled samples to train a suitable network for the present task. Large data distribution differences of gearbox signals are another significant challenge affecting the intelligent fault diagnosis model performance in varying working conditions.

The proposed method

In this study, researchers proposed a gearbox fault diagnosis method based on TL and a lightweight channel attention mechanism. The objective was to solve the existing fault diagnosis challenges and accurately classify limited gearbox samples under cross-component and varying working conditions.

A new EfficientNetV2 network-based model was proposed, which employed the channel attention mechanism to optimize the dimension reduction’s negative impact through proper cross channels. EfficientNetV2, a high-precision and lightweight model, was selected as the basic network to reduce the computation cost. This is a new lightweight CNN that combines neural network search technology.

Initially, the time-frequency distribution of original signals was obtained using wavelet transform to reflect the signals’ local characteristics intuitively. Then, a lightweight, efficient channel attention (LECA) mechanism was designed based on a local cross-channel interaction strategy.

Multi-scale feature input was utilized to retain more detailed features of various dimensions. Channel coefficients and numbers affected the kernel size of one-dimensional (1D) convolution, and a lightweight CNN was constructed. Specifically, the cross-channel size impacted by channel coefficients and the number was adjusted depending on a local cross-channel interaction strategy without any dimensionality reduction, which made the attention mechanism lightweight.

Eventually, a TL method was applied to fine-tune higher structures of the model and freeze lower structures of the network using small samples. This method was used to realize fault diagnosis with high precision for small samples of untrained working components and conditions.

Experimental evaluation and findings

The gearbox dataset obtained from Southeast University was utilized as the fault diagnosis experimental data. The data was obtained on a drivetrain dynamic simulator (DDS), which consists of a two-stage parallel shaft spur gearbox, a two-stage planetary gearbox, a variable speed drive motor, and a programmable controller.

This dataset contained two sub-datasets, including gearbox and bearing datasets, each having four fault states and one health state. Researchers investigated the fault diagnosis effectiveness based on attention mechanisms by comparing and analyzing the classification accuracy of the proposed EfficientNetV2-LECA, EfficientNetV2-efficient channel attention (EfficientNetV2-ECA), and EfficientNetV2-SE models.

They also evaluated the TL fault diagnosis performance of the LECA-EfficientNetV2 network by comparing its classification accuracy with seven other TL fault diagnosis models, including Vgg13, ResNet50, MobileNetV3-L, EfficientNet-b0, EfficientNetV2-S, GhostNetV2, and FasterNet-T2, in four tasks.

Experiment results demonstrated that the proposed LECA-EfficientNetV2 model had the highest fault diagnosis accuracy on both bearing and gear samples. LECA-EfficientNetV2 achieved 99.75% and 99.38% accuracy on the gear and bearing samples, respectively. ECA-EfficientNetV2 attained the second-highest accuracy.

Additionally, LECA-EfficientNetV2 displayed the lowest fault diagnosis time of 13.57 s on bearing samples and 13.22 s on gear samples. These results indicated that LECA-EfficientNetV2 possessed the best diagnostic effect on both datasets and could complete gearbox fault diagnosis more effectively and accurately when compared with ECA-EfficientNetV2 and SE-EfficientNetV2.

Results from the TL experiments showed that LECA-EfficientNetV2 had the best generalization ability and diagnostic performance under various gearbox working conditions and components. The classification accuracy of the proposed model was over 99% in all four tasks and the highest among all eight TL fault diagnosis models evaluated in this study.

To summarize, the findings of this study demonstrated that the proposed LECA-EfficientNetV2 model had robust generalization ability and could be used to realize good fault diagnosis performance under limited samples.