Supervised Health Stage Prediction Using Convolutional Neural Networks for Bearing Wear

Sungho Suh, Joel Jang, Seungjae Won, Mayank Shekhar Jha, Yong Oh LeeSensors 2020 · DOI

TL;DR

This paper introduces a novel method for predicting the health stage of rotating machinery, specifically focusing on bearing wear. By transforming vibration data into images and using Convolutional Neural Networks (CNNs), the approach effectively identifies early signs of degradation without requiring manual threshold setting, outperforming traditional methods in early fault detection.

What Problem Does This Paper Address?

Predicting the health stage (HS) of machinery is crucial for preventing catastrophic failures, ensuring safety, and optimizing maintenance. Current methods often struggle with:

  • Inefficient Feature Extraction: Difficulty in identifying subtle early signs of wear from complex vibration data.
  • Manual Thresholding: Reliance on predefined thresholds for health indicators (HIs), which are sensitive to operational conditions and can lead to high false alarm rates or missed detections.
  • Expert Domain Knowledge: The need for specialized knowledge to set thresholds and interpret data, which is costly and time-consuming.
  • Labeling Effort: Supervised methods typically require extensive labeled data, which is laborious and prone to human error.

This paper aims to overcome these limitations by proposing a data-driven, supervised approach that automates feature extraction and eliminates the need for manual thresholding.

Methods

The proposed method, termed CNN-HS, involves two main phases:

  1. Image Transformation (Nested Scatter Plot - NSP): Raw multi-channel vibration time series data is transformed into a fixed-size image. This involves:
  2. Feature Extraction: Applying bandpass filters to the vibration signals.
  3. Decomposition: Decomposing signals within different bandwidths into nested clusters.
  4. Aggregation: Mapping signals to x- and y-axes and coloring them red, green, and blue to create scatter plots, which are then aggregated into a single RGB image. This process aims to capture multi-variable correlations and reduce non-stationarity.
  5. Health Stage Prediction (Convolutional Neural Network - CNN): The generated NSP images are used as input to a CNN model.
  6. CNN Architecture: A CNN with three convolutional layers and two fully connected layers is employed.
  7. Supervised Learning: The model is trained in a supervised manner using a binary classification approach. It distinguishes between "healthy" (early stage) and "unhealthy" (late stage) data. Crucially, only a small amount of labeled data is required, with the "healthy" label applied to the initial portion of the degradation data and "unhealthy" to the final portion.
  8. Trigger Mechanism: Instead of a threshold, a simple continuous trigger mechanism is used. The machinery is classified as unhealthy when the model outputs a certain number of consecutive "unhealthy" predictions. This helps prevent oscillations and uncertainty.

Datasets

The proposed method was evaluated on two widely used, publicly available datasets:

  • IMS Dataset: Acquired from the NSF I/UCR Center for Intelligent Maintenance Systems (IMS). This dataset consists of run-to-failure experiments on bearings. The study utilized the first set, specifically bearing 4, which exhibited a roller element defect.
  • FEMTO Dataset: Collected from the PRONOSTIA platform for accelerated bearing degradation tests. This dataset comprises 17 run-to-failure experiments under varying operating conditions (rotation speed and radial load).

The paper does not introduce a new benchmark dataset but uses these existing ones to validate its approach against established methods.

Key Insights and Ideas

  • NSP as an Effective Feature Extractor: The NSP method successfully transforms complex time-series vibration data into a visual representation (images) that captures multi-variable correlations and subtle degradation patterns, making them suitable for CNN analysis.
  • Eliminating Manual Thresholds: The CNN-HS model, by learning the distinction between healthy and unhealthy states directly from data and using a trigger mechanism, completely removes the need for manual threshold setting, a major bottleneck in traditional methods.
  • Early Fault Detection: The CNN-HS model demonstrates a superior ability to detect early symptoms of bearing wear compared to traditional RMS and AE-based methods, providing a significantly earlier First Predicting Time (FPT).
  • Simplified Labeling: The binary classification approach (healthy vs. unhealthy) and the focus on the initial and final stages of degradation drastically reduce the labeling effort required for supervised learning.

Key Results

  • Faster Fault Detection: On average, the CNN-HS method predicted faults 47.65% faster than RMS and 44.80% faster than AE methods on the FEMTO dataset.
  • Clearer Health Stage Division: The feature maps generated by the CNN-HS model showed a clearer distinction between healthy and unhealthy states compared to RMS and AE, indicating better HS division.
  • Reduced Ambiguity: The CNN-HS method exhibited less ambiguity in FPT determination compared to RMS and AE.
  • Low Prediction Error for RMS: When used to predict RMS values, the CNN-HS model achieved a low Root Mean Square Error (RMSE), indicating its capability in capturing degradation patterns.

Limitations and Future Work

  • Data Requirements: The method, being supervised, still requires a sufficient amount of training data, which can be challenging to collect in real industrial scenarios where degradation data is scarce.
  • Future Work:
    • More Granular Health Stages: The authors plan to divide the health state into more stages beyond just "healthy" and "unhealthy" to provide a more nuanced health indicator for Remaining Useful Life (RUL) prediction.
  • Unsupervised NSP Application: Extending the use of NSP images in an unsupervised manner for condition-independent fault detection.
  • Testing with Normal Data: Future experiments will include testing the proposed method on a test bench with both degrading bearing data and normal (non-degrading) bearing data to further assess its robustness.

Summary updated

BibTeX
@article{suh2020supervised,
  author = {Sungho Suh and Joel Jang and Seungjae Won and Mayank Shekhar Jha and Yong Oh Lee},
  title = {{Supervised Health Stage Prediction Using Convolutional Neural Networks for Bearing Wear}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20205846},
}
Figure