Distribution-agnostic Probabilistic Few-shot Learning for Multimodal Recognition and Prediction | JoVE Visualize

Area of Science:

Prognostics and Health Management (PHM)
Machine Learning
Reliability Engineering

Background:

Industrial systems require accurate failure mode recognition and Remaining Useful Lifetime (RUL) prediction for effective Prognostics and Health Management (PHM).
Insufficient sensor data is a common challenge in industrial settings, hindering traditional machine learning approaches.
Existing Few-Shot Learning (FSL) methods often treat failure mode recognition and RUL prediction as separate problems, failing to account for their inherent interdependence.

Purpose of the Study:

To develop a novel distribution-agnostic probabilistic FSL method that jointly addresses failure mode recognition and RUL prediction.
To capture the complex dependence between different failure modes and their impact on the RUL of operating units.
To enhance the accuracy and robustness of PHM in industrial scenarios with limited sensor data.

Main Methods:

Proposed a neural network with prototypes to integrate few-shot classification and regression for multimodal recognition and prediction.
Developed multimodal Bayesian model-agnostic meta-learning (MBMAML) to probabilistically model failure modes and RUL under data scarcity.
Constructed a loss function based on probabilistic modeling to train the model, capturing the interaction between failure modes and RUL.

Main Results:

The proposed model adaptively learns approximate distributions of failure modes and RUL for new operating units.
Demonstrated effective performance in a case study involving the degradation of aircraft gas turbine engines.
The method successfully addresses the challenge of limited sensor data by jointly learning failure modes and RUL.

Conclusions:

The developed probabilistic FSL method offers a robust solution for joint failure mode recognition and RUL prediction in data-scarce industrial environments.
MBMAML provides a powerful framework for capturing uncertainty and interdependencies crucial for accurate PHM.
The approach is adaptable and shows promise for practical applications in various industrial settings.