Fault Diagnosis of Rotating Machinery Based on Improved Self-Supervised Learning Method and Very Few Labeled Samples

Meirong Wei, Yan Liu, Tao Zhang, Ze Wang, Jiaming Zhu2022 · DOI

TL;DR

This paper introduces DTC-SimCLR, a novel fault diagnosis method for rotating machinery that excels with very few labeled samples. By combining data transformations with self-supervised learning (SimCLR), it significantly improves diagnostic accuracy and reduces the risk of overfitting, making intelligent fault diagnosis more accessible in real-world industrial scenarios.

What Problem Does This Paper Address?

Rotating machinery, like wind turbines and milling machines, are critical in industrial applications. Their reliable operation is paramount for safety and economic efficiency, necessitating robust fault diagnosis systems. Traditional intelligent fault diagnosis methods, especially those relying on deep learning like Convolutional Neural Networks (CNNs), demand a large number of labeled samples for effective training. However, acquiring such extensive labeled datasets in real-world industrial settings is often difficult, costly, or even impossible. This scarcity of labeled data leads to a significant challenge: overfitting, where models perform well on training data but poorly on unseen data. This paper directly addresses this "very few labeled samples" problem, aiming to develop a fault diagnosis method that is both accurate and data-efficient.

Methods

The core of the proposed method, DTC-SimCLR, lies in two key innovations:

  • Data Transformation Combination (DTC): Instead of traditional data augmentation techniques that artificially increase the dataset size, DTC focuses on enhancing the complexity of existing samples. It applies a combination of transformations (like rotation, horizontal flip, grayscale, cropping, color jitter, and affine) to the original signal data. The selection and parameters of these transformations are guided by mutual information (MI) to maximize the difference between the original and transformed samples. Crucially, this process doesn't increase the number of training samples, thus avoiding the computational burden of generating new data. The transformed data is then reshaped into a 2D matrix and flattened into a 1D sample for processing.
  • Improved Self-Supervised Learning (1-D SimCLR): The paper adapts the SimCLR framework for 1D signal data. This involves:
  • Data Transformation: Creating two correlated "views" of each data sample using the selected DTC. These form a "positive pair."
    • Feature Extraction: A ResNet-based feature encoder (f(·)) extracts representative features from these views.
  • Non-linear Mapping: A fully-connected neural network (g(·)) further processes these features.
  • Contrastive Loss: A contrastive loss function is employed to pull representations of positive pairs closer together in the feature space while pushing representations of negative pairs (from different original samples) further apart.

The DTC-SimCLR model is then constructed by combining the trained feature encoder with a fully-connected layer classifier. During the final fault diagnosis stage, the parameters of the feature encoder are fixed, and only the classifier is trained using the very few available labeled samples. This approach leverages the powerful representations learned during self-supervised training to achieve high accuracy with minimal labeled data.

Datasets

The proposed method was evaluated on two distinct datasets:

  1. Cutting Tooth Fault Dataset: This dataset was collected from a high-speed CNC milling machine. It includes vibration signals from the workpiece and accelerometers measuring forces and vibrations in three directions. The data represents five fault stages: initial wear (IW), mild wear (MW), rapid wear (RW), severe wear (SW), and complete wear (CW). The paper utilized a specific cutter tooth record (c4) and created datasets with 1% of total samples per dataset (10 samples/dataset) as labeled samples, with 300 samples per condition for testing.
  2. Bearing Fault Dataset: This dataset was collected from a test rig with an electric motor, shaft, rolling bearing, flywheel, and load motor. It includes condition signals from testing bearings under three conditions: normal, inner race fault (IRF), and outer race fault (ORF). Similar to the cutting tooth dataset, they created labeled datasets with 1% of total samples per condition (10 samples/dataset) and 300 samples per condition for testing.

Both datasets were uniformly converted into 2D data samples of size 32x32.

Key Insights and Ideas

  • DTC as a Complexity Enhancer, Not a Data Multiplier: The paper's novel approach to data transformation is not about creating more data but about making the existing data more challenging for the model to learn from, thereby mitigating overfitting.
  • Mutual Information for Optimal Transformation Selection: Using MI to guide the selection and parameterization of data transformations is a principled way to ensure these transformations effectively increase sample complexity.
  • Self-Supervised Learning for Robust Feature Extraction: SimCLR, adapted for 1D signals, proves highly effective in learning generalizable feature representations from unlabeled data, which are crucial when labeled data is scarce.
  • Decoupling Feature Learning and Classification: By fixing the feature encoder trained via self-supervision and only training the classifier on few labeled samples, the model efficiently leverages prior knowledge and minimizes the need for extensive labeled data.
  • ResNet as a Powerful Feature Encoder: The choice of ResNet as the backbone for the feature encoder demonstrates its suitability for extracting complex patterns from time-series data.

Key Results

The DTC-SimCLR method demonstrated superior performance across both datasets, especially under conditions of very few labeled samples:

  • Cutting Tooth Fault Diagnosis:
    • Achieved an accuracy of 93.4% with the normalization DTC alone.
  • When comparing DTC-SimCLR against CNN without DTC and CNN with DTC, DTC-SimCLR achieved the highest accuracy and lowest loss with few epochs.
  • t-SNE visualizations showed that DTC-SimCLR produced more separable learned features compared to other methods.
  • In trials with 50 labeled samples, DTC-SimCLR achieved a mean testing accuracy of 93.11% ± 0.24%. This performance was comparable to a CNN trained with 3500 labeled samples.
  • Bearing Fault Diagnosis:
    • The average testing accuracy of DTC-SimCLR was 83.52% across ten trials, outperforming CNN (67.32%) and CNN with DTC (79.24%).
  • The t-SNE visualizations for bearing data also indicated that DTC-SimCLR learned more effective information and good representation from the signals.

Limitations and Future Work

  • Limitations:
    • The paper focuses on 1D signal data. While effective, extending this to multi-channel or multi-sensor data might require further adaptation.
  • The computational cost of training the self-supervised encoder, while less than supervised training with large datasets, is still a factor.
  • Future Work:
    • The authors suggest that the proposed DTC-SimCLR method has the potential to achieve transfer learning between different machine objects, as it only requires a few labeled samples. This is a significant avenue for future research, aiming to build more generalized fault diagnosis systems.

Summary updated

BibTeX
@misc{wei2022fault,
  author = {Meirong Wei and Yan Liu and Tao Zhang and Ze Wang and Jiaming Zhu},
  title = {{Fault Diagnosis of Rotating Machinery Based on Improved Self-Supervised Learning Method and Very Few Labeled Samples}},
  year = {2022},
  doi = {10.3390/s22010192},
}
Figure