A Restricted-Learning Network with Observation Credibility Inference for Few-Shot Degradation Modeling

Academic Background

In complex engineering systems, multiple sensors are widely used to monitor the degradation processes of equipment and predict their Remaining Useful Life (RUL). However, ensuring predictive performance remains challenging when only a limited number of samples are available. Under few-shot scenarios, discordant observations in sensor data introduce significant uncertainty, causing the empirical loss to deviate far from the expected loss. Additionally, the learned degradation models tend to overfit on the limited available samples, leading to biased model parameter distributions, which restrict the model’s generalization ability on unseen data. To address these issues, this paper proposes a restricted-learning network with Observation Credibility Inference (OCI) for few-shot degradation modeling.

This research aims to develop a degradation modeling method suitable for few-shot scenarios, enabling condition monitoring and RUL prediction. It addresses two practical challenges: identifying and handling discordant observations in sensor data and preventing overfitting in few-shot scenarios.

Source of the Paper

This paper is co-authored by Ying Wang, Fangyu Li, Di Wang, and Wei Qin, affiliated with the Department of Industrial Engineering and Management, School of Mechanical Engineering, Shanghai Jiao Tong University, and the School of Information Science and Technology, Beijing University of Technology. The paper was accepted on November 24, 2024, and published in the IEEE Transactions on Automation Science and Engineering. The research was supported by the National Key Research and Development Program of China, the National Natural Science Foundation of China, and the Natural Science Foundation of Shanghai.

Research Process

1. Observation Credibility Inference (OCI)

To identify discordant observations in sensor data, the authors propose the OCI method. By analyzing sensor data, OCI can detect observations that deviate from expected behaviors. These discordant observations, typically caused by measurement errors, transmission errors, or sensor failures, significantly increase model uncertainty. OCI introduces an incident parameter to compensate for the Gaussian-distributed noise term, thereby characterizing the abnormal deviations of discordant observations.

2. Restricted-Learning Network

OCI is integrated into the restricted-learning network. Restricted learning enforces a prior distribution constraint on the degradation model parameters by deleting discordant observations from sensor data, thereby preventing overfitting. This approach ensures that the model parameter distribution approximates a predefined target prior distribution, improving the model’s generalization ability on unseen data.

3. Posterior-Augmented Classifier

After removing discordant observations, the authors construct a posterior-augmented classifier to estimate the health status based on posterior sensor paths and further predict RUL. This classifier is optimized by minimizing the expected loss with respect to the posterior sensor paths, ensuring reliable predictions even with limited samples.

Key Results and Logical Relationships

1. Effectiveness of OCI: OCI successfully identifies discordant observations in sensor data. By deleting or correcting these observations, it significantly improves the accuracy of degradation model parameter estimation.
2. Effect of Restricted Learning: Through restricted learning, the model parameter distribution is effectively constrained, preventing overfitting and enhancing the model’s generalization ability on unseen data.
3. Performance of Posterior-Augmented Classifier: Based on corrected sensor data, the posterior-augmented classifier accurately predicts equipment health status and RUL. In a case study using aircraft engine degradation datasets, the method outperforms benchmark approaches under few-shot scenarios.

Conclusions and Significance

This study proposes a restricted-learning network incorporating OCI, effectively addressing the two major challenges in few-shot degradation modeling: handling discordant observations and preventing model overfitting. By introducing OCI to identify and correct discordant observations and applying restricted learning to enforce prior distribution constraints on model parameters, the method significantly improves prediction performance and generalization ability. This research provides a novel solution for degradation modeling and RUL prediction in engineering systems, particularly valuable in data-scarce scenarios.

Research Highlights

1. Innovative OCI Method: By introducing an incident parameter, OCI effectively identifies and handles discordant observations in sensor data, significantly improving data quality.
2. Introducing Restricted Learning: By enforcing the model parameter distribution to approximate a predefined prior distribution, restricted learning effectively prevents overfitting in few-shot scenarios.
3. Design of the Posterior-Augmented Classifier: This classifier estimates health status based on posterior sensor paths, providing a more reliable basis for RUL prediction.
4. Practical Application Value: This research offers an efficient method for equipment condition monitoring and RUL prediction in engineering systems, with broad application prospects, especially in data-scarce scenarios.

Additional Valuable Information

The authors have publicly released the relevant code and data, facilitating reproducibility and further research by other scholars. This study provides significant theoretical and practical references for the application of few-shot learning in the field of degradation modeling.