Rotating Machine Condition Monitoring for Unlabeled Fault Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data-driven predictive maintenance models for machinery face challenges in learning from unlabeled sensing observations, adapting to novel conditions not represented in initial training data, and handling shifts in data due to emerging faults, as they often rely on supervised learning which struggles with compact and discriminative feature representation in scenarios with limited classes and unlabeled data.
Innovation Solution
The proposed solution involves using Barlow Twins Self-Supervised Learning with Mixed-Up Experience Replay (MixER) analysis, which applies augmentation transformations to 1D time series data from rotating machinery, and Federated Learning to enhance feature generality and adaptability, allowing the model to learn from unlabeled data and adapt to new conditions without ground truth labels, and shares knowledge across machines to improve generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used for predictive maintenance models, then the model can learn from labeled training data, but it struggles to generalize to unseen target domains with limited classes and unlabeled data
Solution Approach 1:
The patent applies self-supervised learning where the model learns to predict augmented versions of its own input data without external labels. The Barlow Twins architecture uses two identical encoders that process augmented views of the same input, and the system learns by minimizing redundancy between their representations, enabling the model to serve its own learning needs without labeled data from unseen domains
Solution Approach 2:
The patent employs data augmentation techniques that transform input data by changing parameters such as time shifts, scaling, and other transformations. These parameter changes create diverse training examples from limited labeled data, improving the model's ability to generalize to unseen domains while maintaining accurate feature representation
2Productivity
If traditional deep learning models are deployed, then they can process CM data, but they assume test-time conditions match training conditions which is often not the case
Solution Approach 1:
The patent creates universal representations through self-supervised learning that are not tied to specific labeled classes or training conditions. The Barlow Twins architecture learns invariant features that function across diverse operating conditions and fault types, making the model universally applicable to various maintenance scenarios without retraining
Solution Approach 2:
The patent implements dynamic adaptation through continual learning mechanisms that allow the model to update its representations as new data becomes available. The system can adapt to changing operating conditions and emerging fault patterns in real-time, maintaining reliability as conditions shift from training to deployment
3Device complexity
If models are trained on limited labeled data, then training is feasible, but the models cannot learn compact discriminative representations for unseen faults
Solution Approach 1:
The model generates its own training signals by comparing augmented views of the same data through two encoders. This self-supervised mechanism creates unlimited pseudo-labels from limited labeled data, enabling the learning of compact discriminative representations without requiring extensive labeled datasets
Solution Approach 2:
The patent performs preliminary self-supervised pretraining on available data before fine-tuning or deployment. This preliminary action of learning robust representations from augmented data prepares the model to handle unseen faults effectively, even when initial labeled data is limited
Data Source
AI summary
The present disclosure describes methods for machine condition monitoring from unlabeled sensing data. The methods include providing a plurality of sensors including at least one acceleration sensor. The plurality of sensors are adapted for obtaining a plurality of high-frequency time series inputs from an operating rotating machine or machine component. The plurality of high-frequency time series inputs received from the plurality of sensors are subjected to one or more augmentation transformations that diversify the plurality of high-frequency time series inputs without impacting condition information. This provides a plurality of augmented inputs which, by the one or more augmentation transformations, are selected to randomize input amplitude and phase.