Color Recurrence Plots for Bearing Fault Diagnosis

Vilma Petrauskiene, Mayur Pal, Maosen Cao, Jie Wang, Minvydas RagulskisSensors 2022 · DOI

TL;DR

This paper introduces "Color Recurrence Plots" (CRPs) as a novel feature extraction technique for diagnosing bearing faults in rotating machinery. By transforming vibration signals into colorful 2D images using non-uniform time lags and averaging, CRPs enhance visual differences between fault types, leading to highly accurate fault classification when combined with deep learning models like AlexNet and SqueezeNet.

What Problem Does This Paper Address?

Rotating machinery, crucial in heavy industries, relies on bearings that are prone to wear and tear, leading to faults. Early detection of these faults is vital to prevent catastrophic failures, downtime, and potential injuries. However, non-intrusive inspection of such machinery is challenging due to inaccessibility. While vibration analysis combined with machine learning has shown promise, existing methods often rely on uniform time delays for feature extraction, potentially missing complex dynamics. This paper addresses the gap in utilizing non-uniform time delays and advanced visualization techniques for more effective bearing fault diagnosis.

Methods

The core innovation is the Color Recurrence Plot (CRP). This method builds upon standard recurrence plots (RPs) but introduces several key advancements:

  • Non-uniform Time-Delay Embedding: Instead of fixed time lags, CRPs use variable time lags (τ₁, τ₂, ..., τ<0xE1><0xB5><0xA2>₋₁) to reconstruct the system's phase space. This is achieved by maximizing a geometric approach-based target function to find optimal time lags.
  • Multi-dimensional Projection and Averaging: The d-dimensional embedded attractor is projected onto multiple 2D planes. For each projection, a recurrence plot is generated.
  • Color Representation: Unlike traditional RPs which are binary (black and white), CRPs assign colors based on the arithmetic average of these projected recurrence plots. This colorization enhances the visual distinction between different fault patterns.
  • Deep Learning Classification: The generated CRP images are then fed into pre-trained deep learning models (AlexNet and SqueezeNet) using a transfer learning approach for image classification. This leverages the powerful feature extraction capabilities of these networks.

Datasets

The study utilizes the Case Western Reserve University Bearing Data Center database. This is a standard benchmark dataset for bearing fault diagnosis research. The data includes vibration signals from bearings with four health states: normal, inner race fault, outer race fault, and ball fault. The authors specifically used data corresponding to a 0.007-inch fault size and a motor speed of 0 HP to demonstrate the sensitivity of their approach to subtle faults.

Key Insights and Ideas

  • Color Enhances Discriminability: The introduction of color in recurrence plots, derived from non-uniform time lags and averaging, significantly improves the visual separation of different bearing fault types compared to traditional black-and-white RPs.
  • Non-uniformity Captures Complex Dynamics: Using variable time lags in the embedding process allows for a more faithful reconstruction of the system's dynamics, which is crucial for accurately representing complex mechanical behaviors.
  • Transfer Learning is Efficient and Effective: Leveraging pre-trained deep learning models like AlexNet and SqueezeNet for image classification of CRPs bypasses the need for designing complex networks from scratch, saving time and computational resources while achieving high accuracy.
  • Data Augmentation and Windowing are Crucial: The study highlights that using overlapping windows for generating averaged recurrence plots, rather than non-overlapping ones, significantly increases the dataset size and improves classification accuracy.

Key Results

The proposed Color Recurrence Plot approach, combined with transfer learning using AlexNet and SqueezeNet, achieved very high classification accuracies:

  • Drive End Bearing Subset:
    • AlexNet: 99.7% accuracy
    • SqueezeNet: 100% accuracy
  • Fan End Bearing Subset:
    • AlexNet: 99.6% accuracy
    • SqueezeNet: 98.4% accuracy
  • Combined Classification (10 classes):
    • Drive End: AlexNet (99.4%), SqueezeNet (98.4%)
    • Fan End: AlexNet (93%), SqueezeNet (95%)

These results demonstrate the efficacy of CRPs in extracting discriminative features for bearing fault diagnosis.

Limitations and Future Work

  • Averaging Method: The paper primarily uses simple arithmetic averaging for generating CRPs. The authors suggest exploring other averaging methods (e.g., harmonic, power law) and comparing their performance.
  • Projection Selection: Currently, all possible planar projections are averaged. Future work could investigate averaging only specific projections or creating multiple groups of CRPs to potentially capture more nuanced information.
  • Dataset Size and Classes: While the study explored both a smaller subset (4 classes) and a larger combined dataset (10 classes), further exploration of intermediate class groupings and their impact on accuracy is suggested.
  • Transfer Learning Nuances: The study focused on AlexNet and SqueezeNet due to their speed. Further research could explore other deep learning architectures and their performance with CRPs.

Summary updated

Abstract

This paper presents bearing fault diagnosis using the image classification of different fault patterns. Feature extraction for image classification is carried out using a novel approach of Color recurrence plots, which is presented for the first time. Color recurrence plots are created using non-linear embedding of the vibration signals into delay coordinate space with variable time lags. Deep learning-based image classification is then performed by building the database of the extracted features of the bearing vibration signals in the form of Color recurrence plots. A Series of computational experiments are performed to compare the accuracy of bearing fault classification using Color recurrence plots. The standard bearing vibration dataset of Case Western Reserve University is used for those purposes. The paper demonstrates the efficacy and the accuracy of a new and unique approach of scalar time series extraction into two-dimensional Color recurrence plots for bearing fault diagnosis.
BibTeX
@inproceedings{petrauskiene2022color,
  author = {Vilma Petrauskiene and Mayur Pal and Maosen Cao and Jie Wang and Minvydas Ragulskis},
  title = {{Color Recurrence Plots for Bearing Fault Diagnosis}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22228870},
}
Figure