An Early Fault Detection Method of Rotating Machines Based on Multiple Feature Fusion with Stacking Architecture

Di WuarXiv 2022 · DOI · arXiv

TL;DR

This paper introduces M2FSA, a novel early fault detection method for rotating machines that fuses multiple types of vibration-based features (time, frequency, and time-frequency domains) using a stacking ensemble architecture. By leveraging deep learning for feature extraction and a hierarchical model, M2FSA achieves more robust and reliable early fault detection with fewer false alarms compared to existing methods.

What Problem Does This Paper Address?

Rotating machines are critical components in many industries, and their failure can lead to catastrophic accidents and significant maintenance costs. Early Fault Detection (EFD) is crucial for implementing condition-based maintenance and preventing such issues. However, existing EFD methods often rely on a single type of feature, which may not capture the full complexity of a machine's health state, especially in the early stages of a fault when the signals are weak and difficult to discern. This can lead to missed detections or an abundance of false alarms. The paper aims to address this by developing a more comprehensive and robust EFD approach that integrates diverse feature information.

Methods

The core of the proposed method, M2FSA (Multiple Feature Fusion with Stacking Architecture), lies in its two-pronged approach:

  1. Multi-Domain Feature Extraction with Deep Learning: The method extracts features from three domains:
  2. Time Domain (TD): Statistical features like RMS, kurtosis, and skewness.
  3. Frequency Domain (FD): Extracted using Fast Fourier Transform (FFT).
  4. Time-Frequency Domain (TFD): Extracted using Hilbert-Huang Transform (HHT). To handle the high dimensionality and noise in FD and TFD features, and to create a unified representation, Stacked Denoising Autoencoders (SDAEs) are employed. Each domain's features are fed into a separate SDAE to learn deep, discriminative representations.

  5. Stacking Ensemble Learning: A two-layer stacking architecture is used for fault detection:

  6. Base Models (Layer 1): Three One-Class Support Vector Machines (OC-SVMs) are trained independently, each taking the deep features from one of the domains (TD, FD, or TFD) as input. OC-SVM is chosen because it can learn a decision boundary using only normal data, which is practical for EFD.
  7. Meta Model (Layer 2): The outputs of the base models (specifically, the distances to their respective hyperplanes) are concatenated and fed into a final OC-SVM meta-model. This meta-model learns to combine the predictions of the base models to make a final decision about the equipment's health state.

The training process utilizes K-fold cross-validation for the base models to prevent overfitting. For early fault alarm, a robust strategy is employed: an alarm is triggered only when a certain proportion of consecutive samples are flagged as abnormal, reducing spurious alarms.

Datasets

The proposed method was evaluated on three publicly available run-to-failure bearing datasets:

  • IEEE PHM Challenge 2012 Dataset: Collected from the PRONOSTIA test platform, featuring vibration and temperature data under three working conditions. The paper specifically uses data from the 1st and 3rd bearings under a specific speed and load.
  • IMS Dataset: This dataset involves degradation experiments on four bearings. The paper focuses on bearing 1 from the second experiment, which developed outer race faults.
  • XJTU-SY Dataset: This dataset includes run-to-failure experiments on 15 rolling bearings under three different working conditions. The paper uses vibration signals from the 5th bearing under the first working condition.

Key Insights and Ideas

  • Synergy of Multiple Features: The paper strongly demonstrates that combining features from different domains (TD, FD, TFD) provides a more comprehensive view of the machine's health than any single domain alone. This is crucial for detecting subtle, early-stage faults.
  • Deep Learning for Robust Representation: SDAEs are effective in extracting low-dimensional, discriminative features from raw vibration signals, helping to overcome noise and dimensionality issues inherent in FD and TFD analysis.
  • Stacking for Enhanced Decision Making: The hierarchical stacking approach, where base models learn from different feature sets and a meta-model combines their insights, leads to a more intelligent and reliable final decision, outperforming individual models.
  • One-Class SVM for Practical EFD: The use of OC-SVM is well-suited for EFD as it requires only normal data for training, making it practical for real-world scenarios where labeled fault data might be scarce.
  • Robust Alarm Strategy: The proposed method's strategy for triggering alarms based on a sequence of abnormal samples significantly reduces false alarms, a common problem in EFD.

Key Results

The M2FSA method demonstrated superior performance across the tested datasets:

  • Earlier Fault Detection: The proposed method consistently detected early faults earlier than most existing methods, often by a significant margin. For example, on the PHM1_1 dataset, it detected the fault at sample 1136, compared to methods detecting it at 1900 or 1690.
  • Reduced False Alarms: A key strength of M2FSA is its significantly lower number of false alarms (FA) compared to other methods, particularly those relying on single feature types or less robust architectures. The paper reports zero false alarms for its method on several test cases.
  • Best Average Score: M2FSA achieved the lowest average score (408.30) across all tested datasets and methods, indicating a superior balance between early fault detection (LO) and false alarms (FA).
  • Feature Visualization: Visualizations of feature spaces (Fig. 5) clearly show that the proposed method's features (Ours) provide better separation between normal, incipient, and severe fault states compared to individual deep domain features or raw/PCA features.

Limitations and Future Work

The authors acknowledge the following limitations and suggest future research directions:

  • Requirement for Normal Data: The proposed method, like other OC-SVM-based approaches, requires a sufficient amount of normal operating data for training the initial models.
  • Fault Type Prediction: The current method focuses on detecting the occurrence of an early fault but does not predict the type of failure.
  • Future Work:
    • Develop a failure type prediction model based on the degradation trends identified by the proposed method.
  • Explore transfer learning or meta-learning techniques to address the "cold start" problem, where limited normal data is available.

Summary updated

BibTeX
@misc{wu2022early,
  author = {Di Wu},
  title = {{An Early Fault Detection Method of Rotating Machines Based on Multiple Feature Fusion with Stacking Architecture}},
  year = {2022},
  eprint = {2205.00511},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.00511},
}
Figure