TL;DR
This paper introduces a novel intelligent fault diagnosis method for hydraulic piston pumps by combining Continuous Wavelet Transform (CWT) for feature extraction with a simplified AlexNet convolutional neural network (CNN) for classification. The proposed method achieves high accuracy and robustness in identifying various fault states, outperforming traditional CNN models.
What Problem Does This Paper Address?
Hydraulic piston pumps are critical components in many industrial and defense systems, but their reliability is challenged by harsh operating conditions. Traditional fault diagnosis methods often rely on expert knowledge and manual feature extraction, which are time-consuming, resource-intensive, and can be prone to errors. The paper addresses the need for an intelligent, automated, and accurate fault diagnosis system for hydraulic piston pumps, particularly by leveraging the power of deep learning for feature learning and classification.
Methods
The core of the proposed method lies in a two-pronged approach:
- Feature Extraction: Continuous Wavelet Transform (CWT) is employed to convert the one-dimensional vibration signals into two-dimensional time-frequency images. CWT is chosen for its ability to effectively handle non-stationary signals, providing high resolution in both time and frequency domains, which is crucial for capturing subtle fault characteristics. These time-frequency images are then processed to create three-channel inputs for the CNN.
- Intelligent Classification: A modified AlexNet architecture is developed. The authors simplify the original AlexNet by reducing the number of convolutional layers from five to three, adjusting the number of kernels, and removing the Local Response Normalization (LRN) layer, which they found to have minor influence on diagnostic accuracy. This simplification aims to reduce computational complexity and the number of parameters, making the model more efficient and less prone to overfitting, especially with limited data. The improved network uses ReLU activation and a softmax classifier for fault classification.
Datasets
The study utilizes a dataset collected from a hydraulic axial piston pump test bench. The dataset comprises vibration signals from five distinct states:
- Normal state
- Sliding slipper wear
- Central spring failure
- Swash plate wear
- Loose slipper
For the failure states (center spring failure, sliding slipper wear, and loose slipper), three levels of severity (minor, moderate, and severe) were introduced to create more comprehensive failure samples. The data was collected under various working pressures (2, 5, 8, 10, and 15 MPa). The dataset was divided into training (70%) and testing (30%) sets.
Key Insights and Ideas
- CWT as a powerful preprocessor: The paper highlights the effectiveness of CWT in transforming complex vibration signals into visually interpretable time-frequency images that effectively capture fault signatures, which are difficult to discern from raw time-domain data.
- Simplified AlexNet for efficiency: By reducing the complexity of the AlexNet architecture, the authors demonstrate that high diagnostic accuracy can be achieved with fewer parameters and computational resources, making the model more practical for real-world applications.
- Integration of feature extraction and deep learning: The synergy between CWT's signal processing capabilities and CNN's deep learning for feature learning and classification is a key enabler of the method's success.
- Robustness to noise and varying conditions: The improved model shows good robustness, even in the presence of noise, which is common in industrial environments.
Key Results
The proposed CWT-AlexNet model achieved a high average recognition accuracy of 98.06% on the test dataset, with the highest accuracy reaching 98.33%. Specifically:
- The normal state, sliding slipper wear, and swash plate wear achieved 100% recognition accuracy.
- Loose slipper failure was recognized with 97.22% accuracy.
- Center spring failure was recognized with 94.72% accuracy.
When compared to traditional models like LeNet-5, improved 2D LeNet-5, and standard AlexNet, the CWT-AlexNet model demonstrated superior performance:
- It achieved the highest average accuracy.
- It exhibited a lower standard deviation, indicating greater consistency.
- It required less training time and was more diagnostically efficient than the classic AlexNet.
Limitations and Future Work
- Data Availability: While the paper presents a comprehensive dataset, the availability of such detailed, labeled data for various fault types and conditions can be a challenge in real-world industrial settings.
- Generalizability: Although the model shows good performance on the collected dataset, further validation on data from different pump models or operating environments would be beneficial to assess its broader generalizability.
- Real-time Implementation: The paper focuses on offline diagnosis. Future work could explore optimizing the model for real-time fault detection and diagnosis in operational systems.
- Explainability: While the t-SNE visualization provides some insight into feature extraction, further research into the interpretability of the deep learning model could enhance trust and understanding of its decision-making process.
Summary updated