Method for Predicting RUL of Rolling Bearings under Different Operating Conditions Based on Transfer Learning and Few Labeled Data

Wei Sun, Haowen Wang, Zicheng Liu, Ronghai QuSensors 2022 · DOI

TL;DR

This paper introduces a novel Domain-Adaptive Adversarial Network (DAAN) for predicting the Remaining Useful Life (RUL) of rolling bearings under varying operating conditions, even with limited labeled data. By leveraging transfer learning and Maximum Mean Discrepancy (MMD), the DAAN effectively bridges the domain gap between different operational states, enabling accurate RUL prediction without extensive retraining for each new condition.

What Problem Does This Paper Address?

Predicting the Remaining Useful Life (RUL) of critical machinery components like rolling bearings is crucial for preventing failures and optimizing maintenance schedules. Traditional deep learning models for RUL prediction face two major hurdles:

  1. Domain Shift: A model trained on data from one operating condition (e.g., a specific load and speed) often fails to generalize to another condition. This is because the underlying data distributions change, rendering the trained model ineffective. Retraining a model for every new operating condition is impractical and data-intensive.
  2. Data Scarcity: Acquiring large amounts of labeled data, especially for fault conditions and their corresponding RUL, is challenging and expensive. This problem is exacerbated when dealing with multiple operating conditions, as labeled data is needed for each.

This paper aims to overcome these limitations by developing a method that can predict RUL across different operating conditions using transfer learning and a minimal amount of labeled data.

Methods

The core of the proposed solution is a Domain-Adaptive Adversarial Network (DAAN). This architecture combines several key techniques:

  • Transfer Learning: This is the fundamental strategy to leverage knowledge gained from a "source domain" (operating condition with sufficient labeled data) to improve performance on a "target domain" (operating condition with limited or no labeled data).
  • Wavelet Packet Transform (WPT): Raw vibration signals are first preprocessed using WPT. This transforms the time-domain signals into a time-frequency representation, creating two-dimensional samples suitable for input into Convolutional Neural Networks (CNNs). WPT is particularly useful for non-stationary signals like vibrations.
  • Convolutional Neural Network (CNN) Feature Extractor (Net1): A CNN is used to extract relevant features from the preprocessed vibration signals. This network is designed to be weight-shared across both source and target domains, promoting domain invariance.
  • Regressor (Net2): This network takes the extracted features and predicts the RUL.
  • Classifier (Net3): This network is trained to distinguish between source and target domain samples.
  • Maximum Mean Discrepancy (MMD): This algorithm is employed to minimize the difference between the feature distributions of the source and target domains. By reducing this discrepancy, the DAAN encourages the feature extractor (Net1) to learn domain-invariant features.
  • Adversarial Training: The DAAN uses an adversarial approach where the classifier (Net3) tries to distinguish between domains, while the feature extractor (Net1) tries to "fool" the classifier by producing domain-invariant features. This is achieved through a combination of classification loss and MMD loss.

The training process involves two main steps that are iteratively repeated:

  1. Step A: The classifier (Net3) is updated to minimize the domain classification loss.
  2. Step B: The feature extractor (Net1) and regressor (Net2) are updated to minimize regression loss (predicting RUL on the source domain), classification loss (adversarial component), and MMD loss (domain adaptation).

Datasets

The experiments utilized a publicly available dataset ([25]) containing vibration signals from rolling bearings. Specifically:

  • Source Domain: Bearing2_1 was used as the source domain in one experimental setup.
  • Target Domains: Bearing3_1 and Bearing3_3 were used as target domains in different experimental cases.
  • Testing: Bearing2_2 was used for testing in one case.

The dataset includes signals from five different load conditions and seventeen whole-bearing remaining life vibration signals. Signals were collected from vertical and horizontal directions. For evaluation, RUL was represented as a percentage of the remaining life.

Key Insights and Ideas

  • Domain Adaptation is Crucial: The paper strongly demonstrates that a model trained on one operating condition cannot reliably predict RUL on another without explicit domain adaptation. This is a fundamental limitation of standard deep learning in real-world industrial settings.
  • MMD as a Bridge: Using MMD to minimize the distribution difference between source and target domain features is a powerful technique for enabling transfer learning in this context. It allows the model to learn representations that are less sensitive to operating condition changes.
  • Adversarial Learning for Invariance: The adversarial component, where the classifier tries to detect the domain and the feature extractor tries to evade detection, is key to learning domain-invariant features. This forces the network to focus on the underlying degradation patterns rather than condition-specific artifacts.
  • Few-Shot Learning for RUL: The method's ability to perform well with limited labeled data in the target domain is a significant practical advantage, directly addressing the data acquisition challenge.

Key Results

  • Improved Accuracy: The proposed DAAN method significantly outperformed methods without transfer learning. For example, in Case I (predicting on bearing3_3), the proposed method achieved an RMSE of 7.01% compared to 27.23% for the non-transfer learning method. Similarly, MAPE was reduced from 48.00% to 18.36%.
  • Effective Domain Transfer: The method demonstrated good generalization across different operating conditions in both tested cases, confirming its ability to transfer learned knowledge.
  • Robustness to Noise: The DAAN showed acceptable performance even when subjected to significant levels of added Gaussian noise (up to a standard deviation of 2g), indicating its robustness in real-world, noisy environments.
  • Comparison with Related Methods: While some specialized methods (like SCAE) showed slightly better performance on specific metrics (e.g., MAPE), they were often limited to predicting only the rapid degradation stage. The proposed DAAN predicts the whole life RUL and achieves good accuracy with less training data compared to some other transfer learning approaches.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the proposed method is more accurate than non-transfer learning approaches, its relative error performance (MAPE) might be slightly poorer in some instances compared to highly specialized methods that focus on specific degradation stages.
  • The amount of source and target domain data used in the proposed method was less than some comparative methods, which could contribute to the observed accuracy differences.
  • Future Work:
    • The paper implicitly suggests that further optimization of the transfer learning strategy and network architecture could lead to even better performance.
  • Exploring more complex operating condition shifts and fault types could be a natural extension.
  • Investigating the impact of different WPT parameters or alternative signal preprocessing techniques might also be beneficial.

Summary updated

Abstract

As industrial development increases, electric machine systems are more widely used in industrial production. Rolling bearings play a key role in machine systems and so the prevention of faults in rolling bearings is more important than ever before. Recently, with the development of artificial intelligence, neural networks have been used to monitor the remaining useful life of rolling bearings. However, there are two problems with this technique. First, a network trained by data for a single operating condition (source domain) cannot predict the remaining useful life of bearings under a different operating condition (target domain), such as a different load or speed. Second, a large number of labeled data are needed for network training, but the acquisition of labeled data for different operating conditions is a challenging task. To address these problems, this paper proposes a domain-adaptive adversarial network, in which a transfer learning strategy and maximum mean discrepancy algorithm are used for network optimization, so that remaining useful life can be predicted without labeled data in target domain training. Our results confirm that a model trained by source domain data alone cannot predict the remaining useful life of bearings under different conditions, but the domain-adaptive adversarial network can accurately predict remaining useful life for varying operating conditions. The method proposed also exhibits good performance even if there are noises in the signals.
BibTeX
@inproceedings{sun2022method,
  author = {Wei Sun and Haowen Wang and Zicheng Liu and Ronghai Qu},
  title = {{Method for Predicting RUL of Rolling Bearings under Different Operating Conditions Based on Transfer Learning and Few Labeled Data}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s23010227},
}
Figure