A Novel Deep Learning Method for Intelligent Fault Diagnosis of Rotating Machinery Based on Improved CNN-SVM and Multichannel Data Fusion

Wenfeng Gong, Hui Chen, Zehui Zhang, Meiling Zhang, Ruihan Wang, Cong Guan, Qin Wang2019 · DOI

TL;DR

This paper introduces an improved CNN-SVM deep learning model for intelligent fault diagnosis in rotating machinery. By replacing fully connected layers with global average pooling and using SVM for classification, the model achieves higher accuracy and significantly reduces computational complexity, making it more suitable for real-time applications.

What Problem Does This Paper Address?

Rotating machinery is crucial in modern industry, and its reliable operation depends on effective fault diagnosis. Identifying incipient micro-faults, which are subtle and easily masked by noise, is particularly challenging. Traditional methods often rely on manual feature extraction by experts, which is time-consuming, subjective, and lacks versatility. While deep learning offers automated feature extraction, existing CNN models can suffer from excessive parameters and long training times due to fully connected layers, hindering real-time diagnosis. This paper aims to address these limitations by proposing a novel, more efficient, and accurate deep learning approach for intelligent fault diagnosis.

Methods

The core of this paper is the Improved CNN-SVM model. It builds upon the standard Convolutional Neural Network (CNN) architecture but introduces key modifications:

  • Global Average Pooling: Instead of traditional fully connected layers, which have a large number of parameters, the model uses global average pooling. This significantly reduces the parameter count and computational load.
  • SVM Classifier: The Softmax classifier, commonly used in CNNs, is replaced with a Support Vector Machine (SVM) for the final classification stage. SVM is known for its strong performance in multi-class classification, especially with complex data.
  • Multichannel Data Fusion: The model directly processes multichannel vibration data from multiple sensors, fusing temporal and spatial information.
  • Deep Learning Training Skills: To combat over-fitting and improve generalization, the model incorporates techniques like batch normalization, data augmentation, and dropout.

The overall process involves:

  1. Training an improved CNN-Softmax model on raw multichannel data.
  2. Using the trained CNN as a feature extractor to obtain representative features from new fault data.
  3. Inputting these extracted features into the SVM classifier for final fault diagnosis.

Datasets

The study utilizes a multichannel fault dataset of rolling bearings collected from the Electrical Engineering Laboratory of Case Western Reserve University (CWRU). This dataset includes vibration signals from drive-end and fan-end bearings under various working conditions (0, 1, and 2 horsepower loads). The dataset comprises 10 health statuses, including normal operation and nine fault types (categorized by fault location and severity). The data was preprocessed, including segmentation into 500-sample segments, standardization, and division into training (70%), verification (20%), and testing (30%) sets.

Key Insights and Ideas

  • Global Average Pooling is a Game Changer: Replacing fully connected layers with global average pooling dramatically reduces model complexity and training time without sacrificing accuracy, making real-time diagnosis feasible.
  • CNN as a Feature Extractor + SVM as Classifier Synergy: Combining the powerful feature extraction capabilities of CNNs with the robust classification of SVMs creates a potent diagnostic system, especially for complex fault patterns.
  • End-to-End Learning is Achievable: The proposed method eliminates the need for manual feature engineering, allowing the model to learn relevant features directly from raw sensor data.
  • Deep Learning Skills are Crucial for Robustness: Techniques like data augmentation and dropout are essential for preventing over-fitting and ensuring the model generalizes well to unseen data, particularly when dealing with limited fault samples.

Key Results

The proposed Improved CNN-SVM model achieved a remarkable test accuracy of 99.94% on the rolling bearing fault dataset. This significantly outperforms traditional methods like SVM (95.01%), KNN (91.15%), BPNN (83.95%), DNN (93.66%), and the standard CNN (98.94%). The model also demonstrated a substantial reduction in test time (0.08318 seconds) compared to traditional CNNs, making it highly suitable for real-time applications. The confusion matrix analysis revealed excellent performance, with only one misclassification out of 1800 test samples, highlighting the model's ability to accurately diagnose even incipient micro-faults.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the model is effective, the training process can still be time-consuming, although significantly reduced compared to traditional CNNs with fully connected layers.
  • The study focuses on rolling bearings; extending the method to other types of rotating machinery would be beneficial.
  • Future Work:
    • The authors suggest exploring the use of CNN-SVM and other deep learning methods for online real-time intelligence diagnosis and hardware implementation.
  • Further research into optimizing the model for different types of rotating machinery and fault scenarios is implied.

Summary updated

BibTeX
@misc{gong2019novel,
  author = {Wenfeng Gong and Hui Chen and Zehui Zhang and Meiling Zhang and Ruihan Wang and Cong Guan and Qin Wang},
  title = {{A Novel Deep Learning Method for Intelligent Fault Diagnosis of Rotating Machinery Based on Improved CNN-SVM and Multichannel Data Fusion}},
  year = {2019},
  doi = {10.3390/s19071693},
}
Figure