Feature Space Transformation for Fault Diagnosis of Rotating Machinery under Different Working Conditions

Gye-Bong Jang, Sung-Bae Cho

TL;DR

This paper tackles the challenge of diagnosing faults in rotating machinery when training and testing data come from different operating conditions (domain shift). They propose a novel "feature space transformation" method that leverages normal data from the target domain to adapt a model trained on a source domain, achieving higher cross-domain diagnostic accuracy.

What Problem Does This Paper Address?

Diagnosing faults in rotating machinery (like gearboxes and bearings) is crucial for preventing costly failures and ensuring safety. While deep learning models have shown promise, a major hurdle in real-world applications is "domain shift." This occurs when a model trained on data from one set of operating conditions (e.g., specific speeds, loads) performs poorly on data from different conditions. Collecting comprehensive failure data across all possible operating scenarios is expensive and time-consuming. This paper aims to bridge this gap by enabling models to generalize well to new, unseen operating conditions using readily available normal data from the target domain.

Methods

The core of the proposed method involves transforming the latent space representation of the data to minimize discrepancies between the source and target domains, while preserving fault-specific information. Key techniques include:

  • Short-Time Fourier Transform (STFT): Used for pre-processing vibration signals to capture time-frequency characteristics, essential for non-stationary mechanical systems.
  • Attentional Autoencoder: An unsupervised learning model that learns a compressed latent representation of the input signal. An attention mechanism is incorporated within the encoder to help the model focus on fault-relevant features.
  • 1D CNN LSTM Classifier: A deep learning architecture designed to process the extracted latent features. 1D CNNs are used to capture local impulse patterns characteristic of mechanical faults, while LSTMs are employed to learn contextual temporal dependencies.
  • Latent Vector Shift for Domain Adaptation: This is the novel contribution. Inspired by word embedding analogies (e.g., "King - Man + Woman ≈ Queen"), they propose a method to shift the latent vectors of the target domain data to align with the distribution of the source domain. This is achieved by using the normal data from the target domain to learn a transformation that effectively "moves" the target domain's latent space closer to the source domain's latent space, without requiring labeled target domain failure data.

Datasets

The proposed model was validated on two types of datasets:

  • Case Study 1: CWRU Dataset: A widely used benchmark dataset for rolling element bearing diagnostics. This dataset contains data from various operating conditions (different speeds and loads) and fault types.
  • Case Study 2: Real Machine Dataset: Data collected from heavy equipment (specifically, a gearbox system) operating under various conditions. This dataset is noted to be more representative of real-world complexities, including noise and varying operating environments.

Key Insights and Ideas

  • Leveraging Normal Data for Domain Adaptation: The paper highlights that collecting normal data from a target domain is significantly easier than collecting failure data. This readily available normal data can be effectively used to adapt models trained on a source domain, a crucial insight for practical deployment.
  • Semantic Transformation in Latent Space: The analogy to word embeddings for transforming latent representations is a powerful conceptual framework. It suggests that semantic relationships learned in one domain can be transferred to another by manipulating their vector representations.
  • Attention for Fault Feature Focus: Incorporating an attention mechanism within the autoencoder helps the model prioritize features that are most indicative of faults, making the latent representation more discriminative.
  • Robustness to Real-World Conditions: The validation on a real machine dataset, with its inherent noise and complex operating environments, demonstrates the practical applicability and robustness of the proposed approach.

Key Results

The proposed model consistently outperformed several state-of-the-art methods across various cross-domain scenarios on both the CWRU and real machine datasets.

  • CWRU Dataset: Achieved an average accuracy of 83.63%, significantly higher than other methods which ranged from 55.24% to 69.11%.
  • Real Machine Dataset: Demonstrated an average accuracy of 82.08%, again outperforming comparative methods.
  • Qualitative Results: t-SNE visualizations showed that the proposed domain transformation effectively brings the latent representations of the source and target domains closer together, even for unlabeled target domain failure data. The model was also successfully deployed on an embedded board and verified on actual excavator operation, accurately identifying existing defects.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the model performs well, there can still be some distance between distributions in certain cases, particularly in the real machine dataset.
  • The CWRU dataset, while a benchmark, is collected in a laboratory environment and may not fully capture the complexities of all real-world scenarios.
  • Future Work:
    • The authors suggest that the proposed system is suitable for practical applications and has great significance in industries like large-scale mining machinery. Further exploration and deployment in such environments would be a natural next step.
  • While the paper focuses on rotating machinery, the underlying domain adaptation principles could potentially be extended to other types of industrial equipment.

Summary updated

Abstract

In recent years, various deep learning models have been developed for the fault diagnosis of rotating machines. However, in practical applications related to fault diagnosis, it is difficult to immediately implement a trained model because the distribution of source data and target domain data have different distributions. Additionally, collecting failure data for various operating conditions is time consuming and expensive. In this paper, we introduce a new transformation method for the latent space between domains using the source domain and normal data of the target domain that can be easily collected. Inspired by semantic transformations in an embedded space in the field of word embedding, discrepancies between the distribution of the source and target domains are minimized by transforming the latent representation space in which fault attributes are preserved. To match the feature area and distribution, spatial attention is applied to learn the latent feature spaces, and the 1D CNN LSTM architecture is implemented to maximize the intra-class classification. The proposed model was validated for two types of rotating machines such as a dataset of rolling bearings as CWRU and a gearbox dataset of heavy machinery. Experimental results show the proposed method has higher cross-domain diagnostic accuracy than others, therefore showing reliable generalization performance in rotating machines operating under various conditions.
BibTeX
@inproceedings{jang2021feature,
  author = {Gye-Bong Jang and Sung-Bae Cho},
  title = {{Feature Space Transformation for Fault Diagnosis of Rotating Machinery under Different Working Conditions}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=jzYJWtqIRMe},
}
Figure