1. Introduction
Gear systems are fundamental components in various mechanical devices, widely used in industries such as automotive, aerospace, and manufacturing. The reliable operation of gear systems is crucial for the overall performance and safety of these applications. However, due to complex working conditions and environmental factors, gears are prone to various faults, which can lead to significant economic losses and even endanger human lives. Therefore, accurate and efficient gear fault diagnosis has become a research hotspot in the field of mechanical engineering.
1.1 The Importance of Gear Fault Diagnosis
Gear failures can cause equipment breakdowns, production interruptions, and increased maintenance costs. In some critical applications, such as aircraft engines and power transmission systems, gear faults may even result in catastrophic accidents. For example, a gear failure in an aircraft engine could lead to a loss of power and potentially a crash. Therefore, early detection and diagnosis of gear faults can help prevent these consequences by enabling timely maintenance and repair actions.
1.2 Traditional Gear Fault Diagnosis Methods
Traditional gear fault diagnosis methods mainly include vibration analysis, oil analysis, and acoustic emission testing. Vibration analysis is the most commonly used method, which involves measuring the vibration signals of the gear system and analyzing the characteristics of the signals to detect faults. For instance, the amplitude and frequency of the vibration signal can change when a gear has a tooth crack or wear. Oil analysis is used to detect the presence of metal particles or other contaminants in the lubricating oil, which can indicate gear wear or damage. Acoustic emission testing measures the high-frequency sound waves emitted by the gear system during operation, and changes in the acoustic emission patterns can suggest the occurrence of faults.
However, these traditional methods have several limitations. They often require extensive signal processing and professional knowledge for feature extraction. The process of manually extracting features from vibration signals, for example, can be time-consuming and inaccurate. Moreover, the diagnosis accuracy and efficiency of these methods are relatively low, especially for complex gear faults. In some cases, early-stage faults may not be detected until they have developed into more severe problems.
2. Deep Learning in Gear Fault Diagnosis
With the development of artificial intelligence and machine learning technologies, deep learning has emerged as a powerful tool for gear fault diagnosis. Deep learning algorithms can automatically learn the features of gear faults from large amounts of data, eliminating the need for manual feature extraction. This has significantly improved the accuracy and efficiency of fault diagnosis.
2.1 Convolutional Neural Networks (CNNs) in Gear Fault Diagnosis
CNNs are a type of deep learning model that has been widely applied in image recognition and signal processing. In gear fault diagnosis, CNNs can be used to analyze vibration signals or other sensor data to detect and classify gear faults. The hierarchical structure of CNNs allows them to learn the local and global features of the data, making them suitable for handling complex fault patterns.
For example, a CNN can be trained to recognize the different vibration patterns associated with healthy gears, gears with tooth wear, and gears with cracks. By analyzing the frequency and amplitude characteristics of the vibration signals, the CNN can accurately classify the gear faults. The advantage of using CNNs is that they can handle large amounts of data and can adapt to different fault scenarios.
2.2 Limitations of Traditional CNNs in Gear Fault Diagnosis
Despite their effectiveness, traditional CNNs have some limitations in gear fault diagnosis. One of the main limitations is the single-channel feature extraction network framework, which results in a relatively low diagnosis accuracy. The use of a single channel may not capture all the relevant features of the gear faults, leading to misclassifications. Another limitation is the large number of fully connected layers in the CNN model, which leads to a high computational cost and low training efficiency. The fully connected layers require a large number of parameters to be trained, which can be time-consuming and may cause overfitting.
3. The Proposed Improved Convolutional Neural Network for Gear Fault Diagnosis
To address the limitations of traditional CNNs, an improved convolutional neural network (IMCNN) is proposed in this study for gear fault diagnosis. The IMCNN incorporates two key improvements: a parallel module structure and a global average pooling layer.
3.1 Parallel Module Structure
The parallel module structure uses different-sized convolution kernels to process the same input layer and then concatenates the outputs. This allows the network to extract multi-scale features from the data, enhancing the feature expression ability. As shown in the following table, the parallel module can capture features at different resolutions, improving the accuracy of fault classification.
| Convolution Kernel Size | Feature Scale |
|---|---|
| 5×1 | Coarse-scale features |
| 3×1 | Medium-scale features |
| 2×1 | Fine-scale features |
By combining these different-scale features, the network can better distinguish between different types of gear faults. For example, a small crack in a gear tooth may be more easily detected by the fine-scale features, while overall wear can be captured by the coarser-scale features.
3.2 Global Average Pooling Layer
The global average pooling layer takes the average of all the data from the previous layer as the output. This significantly reduces the number of trainable parameters compared to the traditional fully connected layers, improving the training efficiency of the network. As shown in the figure below, the global average pooling layer simplifies the network structure and accelerates the convergence of the model.
[Insert a diagram illustrating the global average pooling layer here]
In addition to improving the training speed, the global average pooling layer also helps to prevent overfitting. By reducing the number of parameters, the model becomes less complex and is less likely to fit the training data too closely, thereby improving its generalization ability.
4. Experimental Setup and Results
To evaluate the performance of the proposed IMCNN, experiments were conducted using a gear fault dataset collected from an SQI (Spectra Quest Inc) power transmission fault diagnosis test bench.
4.1 Experimental Setup
The test bench consisted of a motor, a coupling, a planetary gearbox, a parallel shaft gearbox, and a magnetic powder brake. The motor drove the magnetic powder brake through the gearboxes, and the vibration signals were collected using accelerometers. Four different types of sun gears were tested: healthy, missing tooth, broken tooth, and worn gears. The sampling frequency was set to 12,800 Hz, and the motor speed was 17 Hz with a current of 0.2 A.
4.2 Data Preprocessing
The collected vibration signals were preprocessed to form a dataset. The signals from the two sensors were extracted and combined to form a total dataset. The data was then normalized and divided into training and testing sets, with 80% of the data used for training and 20% for testing. Each sample consisted of 154 rows of data, and the dataset was formatted as 21,328×154×2. The gear fault states were labeled as “0” for healthy, “1” for missing tooth, “2” for broken tooth, and “3” for worn gears.
4.3 IMCNN Model Training and Testing
The IMCNN model was trained using the training set, and the performance was evaluated using the testing set. The training process involved forward propagation to extract features and calculate the error, and then backward propagation to adjust the network parameters. The training was repeated until the maximum number of training epochs was reached. The accuracy and training time of the IMCNN were compared with those of the traditional CNN and the MCNN (CNN with parallel module).
4.4 Experimental Results
The results showed that the IMCNN achieved a higher accuracy and shorter training time compared to the traditional CNN and the MCNN. The accuracy of the IMCNN reached 97.28%, which was 3.61% higher than that of the CNN and 0.83% higher than that of the MCNN. The training time of the IMCNN was 1,181 s, which was 449 s shorter than that of the CNN and 236 s shorter than that of the MCNN. The error curve during training showed a decreasing trend, indicating that the model was converging well. The accuracy curve on the testing set also showed an increasing trend and finally stabilized at a high level.
5. Comparison with Other Methods
To further demonstrate the superiority of the proposed IMCNN, a comparison was made with other existing gear fault diagnosis methods.
5.1 Comparison with Traditional Diagnosis Methods
Compared with traditional methods such as vibration analysis and oil analysis, the IMCNN has several advantages. Firstly, the IMCNN can automatically learn the features of gear faults from data, eliminating the need for manual feature extraction. This saves time and reduces the dependence on professional knowledge. Secondly, the IMCNN has a higher diagnosis accuracy and can detect early-stage faults more effectively. Traditional methods may miss some subtle fault features, while the IMCNN can capture these features through its deep learning architecture.
5.2 Comparison with Other Deep Learning Methods
In addition to the comparison with the traditional CNN and the MCNN, the IMCNN was also compared with other deep learning methods such as the deep belief network. The results showed that the IMCNN outperformed these methods in terms of accuracy and efficiency. The IMCNN’s parallel module and global average pooling layer contributed to its better performance, enabling it to handle complex gear fault patterns more accurately and efficiently.
6. Conclusion and Future Work
In this study, an improved convolutional neural network (IMCNN) was proposed for gear fault diagnosis. The IMCNN incorporated a parallel module structure and a global average pooling layer to improve the diagnosis accuracy and efficiency. Experimental results using a gear fault dataset demonstrated the effectiveness of the proposed method, showing higher accuracy and shorter training time compared to traditional CNN and MCNN methods.
6.1 Conclusion
The parallel module structure allowed the network to extract multi-scale features, enhancing the feature expression ability and improving the diagnosis accuracy. The global average pooling layer reduced the number of trainable parameters, improving the training efficiency and preventing overfitting. The proposed IMCNN has significant potential for practical applications in gear fault diagnosis, helping to improve the reliability and safety of mechanical systems.
6.2 Future Work
Despite the good performance of the proposed method, there are still some areas for improvement. Future work will focus on further optimizing the network structure and parameters. For example, the determination of the initial learning rate and the batch size of samples needs to be studied further to improve the learning effect of the network. Additionally, the number of convolution blocks, pooling layers, and neurons in each layer of the network also affects the diagnosis performance, and methods for determining these parameters need to be explored. By addressing these issues, the performance of the gear fault diagnosis model can be further enhanced, providing more accurate and efficient fault diagnosis solutions for industrial applications.
