Fault Detection of Bearing: An Unsupervised Machine Learning Approach Exploiting Feature Extraction and Dimensionality Reduction

Lucas Costa Brito, Gian Antonio Susto, Jorge Nei Brito, Marcus Antonio Viana DuarteInformatics 2021 · DOI · OpenReview

TL;DR

This paper tackles the challenge of detecting bearing faults in rotating machinery using unsupervised machine learning. By exploring various feature extraction and dimensionality reduction techniques, the authors demonstrate that carefully selected features, combined with methods like PCA, can achieve high accuracy in anomaly detection, even with reduced data complexity. This approach offers a more efficient and interpretable way to monitor machinery health.

What Problem Does This Paper Address?

Monitoring the health of rotating machinery is crucial for industrial applications, and bearings are particularly critical components. Traditional methods often rely on specialists to analyze vast amounts of sensor data, which can be time-consuming and prone to errors, especially with the increasing number of monitored equipment. While machine learning (ML) offers promising solutions, the sheer volume of data and the complexity of extracting meaningful information present significant hurdles. Specifically, this paper addresses the challenge of identifying bearing faults (anomalies) in an unsupervised manner, meaning without pre-labeled data of fault conditions. This is important because obtaining such labels in real-world industrial settings is often difficult and expensive. Furthermore, the paper aims to investigate how feature extraction and dimensionality reduction can improve the accuracy and efficiency of these ML models, making them more practical for real-time monitoring.

Methods

The core of this research lies in a five-step methodology:

  1. Data Acquisition: Vibration signals were collected from bearings.
  2. Feature Extraction: A comprehensive set of 22 features was extracted from the vibration signals. These features were categorized into time-domain (e.g., RMS, kurtosis, skewness), frequency-domain (e.g., energy in specific bands), and time-frequency domain (e.g., wavelet entropy). The authors also considered features derived from expert knowledge and signal processing techniques like envelope analysis.
  3. Dimensionality Reduction: Several techniques were employed to reduce the number of features while preserving essential information. These included Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), Isometric Feature Mapping (ISOMAP), Independent Component Analysis (ICA), and Autoencoders (AE). The paper also explored reducing the dimensionality of the raw signal itself.
  4. Fault Detection (Anomaly Detection): The Isolation Forest (IF) algorithm, a popular unsupervised anomaly detection method, was used. IF works by isolating anomalies, which require fewer partitions than normal data points.
  5. Feature Trend Analysis: This step involved visualizing the behavior of extracted features over time to understand their correlation with fault progression and to aid human specialists in interpreting the model's findings.

A key novelty is the systematic comparison of various feature sets (all extracted, manually selected, and dimensionality-reduced) and dimensionality reduction techniques within an unsupervised anomaly detection framework for bearing fault diagnosis.

Datasets

The study utilized a publicly available bearing dataset from the University of Cincinnati's Center for Intelligent Maintenance Systems. This dataset consists of run-to-failure tests with multiple bearings, providing vibration signals. For this study, bearing 01 from test 02 was used, with the data manually labeled for anomaly detection evaluation.

Key Insights and Ideas

  • Feature Engineering is Paramount: The paper strongly emphasizes that the quality and relevance of extracted features are critical for the success of ML-based fault detection. Manually selected features, based on expert knowledge of bearing faults (like BPFO), yielded the best results, highlighting the value of domain expertise.
  • Dimensionality Reduction Offers a Trade-off: While dimensionality reduction techniques like PCA can significantly reduce data complexity and computational cost, they can sometimes lead to a loss of explainability. However, for anomaly detection, PCA proved to be a strong performer, achieving accuracy close to manually selected features.
  • Unsupervised Learning is Viable: The study successfully demonstrates the effectiveness of unsupervised anomaly detection (Isolation Forest) for identifying bearing faults, overcoming the challenge of obtaining labeled fault data.
  • Trend Analysis Aids Interpretation: Visualizing feature trends provides valuable insights for specialists, helping them understand the progression of a fault and complementing the automated anomaly scores.

Key Results

  • Best Performance: The set of manually selected features achieved the highest F1-Score (98.02%), closely followed by PCA-reduced features (97.45%) and ICA-reduced features (97.39%).
  • Dimensionality Reduction Effectiveness: PCA, t-SNE, and ICA demonstrated good results after dimensionality reduction, indicating their ability to represent the data effectively in a lower-dimensional space.
  • Raw Signal Reduction: Reducing the dimensionality of the raw signal using PCA also showed promise, though it required a higher number of principal components to achieve comparable results to extracted features.
  • Confusion Matrix Insights: The confusion matrix revealed a low rate of false negatives, indicating the model's ability to detect most faults, and errors primarily occurred during the incipient stages of the fault.

Limitations and Future Work

  • Loss of Explainability: A noted limitation of dimensionality reduction techniques is the potential loss of interpretability of the reduced features, which can be a drawback in real-world applications where understanding why an anomaly is detected is crucial.
  • Future Work: The authors suggest future research focusing on understanding how the model relates input features to their significance in anomaly detection. This could lead to more optimized dimensionality reduction strategies and potentially contribute to the field of Explainable Artificial Intelligence (XAI) in machinery monitoring. They also aim to explore more optimized ways to perform dimensionality reduction.

Summary updated

Abstract

The monitoring of rotating machinery is an essential activity for asset management today. Due to the large amount of monitored equipment, analyzing all the collected signals/features becomes an arduous task, leading the specialist to rely often on general alarms, which in turn can compromise the accuracy of the diagnosis. In order to make monitoring more intelligent, several machine learning techniques have been proposed to reduce the dimension of the input data and also to analyze it. This paper, therefore, aims to compare the use of vibration features extracted based on machine learning models, expert domain, and other signal processing approaches for identifying bearing faults (anomalies) using machine learning (ML)—in addition to verifying the possibility of reducing the number of monitored features, and consequently the behavior of the model when working with reduced dimensionality of the input data. As vibration analysis is one of the predictive techniques that present better results in the monitoring of rotating machinery, vibration signals from an experimental bearing dataset were used. The proposed features were used as input to an unsupervised anomaly detection model (Isolation Forest) to identify bearing fault. Through the study, it is possible to verify how the ML model behaves in view of the different possibilities of input features used, and their influences on the final result in addition to the possibility of reducing the number of features that are usually monitored by reducing the dimension. In addition to increasing the accuracy of the model when extracting correct features for the application under study, the reduction in dimensionality allows the specialist to monitor in a compact way the various features collected on the equipment.
BibTeX
@article{brito2021fault,
  author = {Lucas Costa Brito and Gian Antonio Susto and Jorge Nei Brito and Marcus Antonio Viana Duarte},
  title = {{Fault Detection of Bearing: An Unsupervised Machine Learning Approach Exploiting Feature Extraction and Dimensionality Reduction}},
  journal = {Informatics},
  year = {2021},
  doi = {10.3390/informatics8040085},
}
Figure