Rotating Machine Monitoring via Domain Adaptation of Unlabelled Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in monitoring rotating parts of rotating machines lies in the difficulty of collecting sufficient labeled data due to the rarity of faults in aerospace, leading to inaccurate deep learning model performance, especially when dealing with unlabelled data from different acquisition contexts.
Innovation Solution
A method utilizing a deep transfer learning model that adjusts parameters independently of target database labels by minimizing a cost function with a Gaussian kernel function based on Pascal's triangle, allowing for domain adaptation between source and target databases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are trained using conventional approaches with labeled data, then classification accuracy can be achieved, but the method fails when applied to unlabelled data from different acquisition contexts due to domain distribution differences
Solution Approach 1:
The patent introduces a domain adapter as an intermediary component between the frozen deep learning model and the unlabelled target domain data. This adapter learns to map features from the source domain distribution to the target domain distribution, enabling the model to adapt to different acquisition contexts without retraining the main model. The adapter processes the feature representations and transforms them to be compatible with the target domain's probability distribution.
Solution Approach 2:
The patent changes the parameters of the domain adapter during training by minimizing the Maximum Mean Discrepancy (MMD) between the source and target domain distributions. The adapter's weights and biases are updated iteratively to reduce the statistical difference between domains. Additionally, the patent dynamically adjusts the Gaussian kernel bandwidth parameter in the MMD calculation based on the training progress, allowing the model to adapt to varying distribution differences.
2Ease of manufacture
If the Gaussian kernel bandwidth is fixed in the MMD calculation, then the computation is simpler, but the model cannot adapt to varying distribution differences between source and target domains
Solution Approach 1:
The patent transforms the static Gaussian kernel bandwidth into a dynamic parameter that evolves during training. The bandwidth is calculated based on the current state of the domain adapter and the distribution characteristics of the data. This allows the kernel to automatically adjust its sensitivity to distribution differences, providing both computational efficiency and adaptability without requiring manual tuning.
3Adaptability or versatility
If domain adaptation is performed without labelled target data, then the method becomes more versatile and applicable to real-world scenarios, but the training process becomes more complex requiring advanced optimization techniques
Solution Approach 1:
The patent replaces complex manual feature engineering and expert knowledge with an automated domain adaptation mechanism. Instead of requiring experts to manually adjust features or create domain-specific models, the system uses the MMD minimization approach to automatically learn the domain transformation. This substitutes the mechanical process of manual tuning with an automated optimization process that handles the complexity internally.
Data Source
AI summary
A method for automatically monitoring a plurality of rotating parts of rotating machines on the basis of a target database including a plurality of time signals from a distribution generated from each rotating part and on the basis of a source database including a plurality of time signals from a distribution S different from the distribution T generated from a source rotating part of a source rotating machine and being associated with an operating class, the monitoring being carried out by an adaptive deep learning model making it possible to adapt the source distribution to the target distribution, the deep learning module being trained by minimization of a cost function relating to Gaussian kernel functions having a parameter σ; σ being calculated in each period on the basis of the difference in distributions weighted by a constant static value estimated on the basis of a Pascal's triangle.