Cross-Domain Prediction Model Transfer for Unlabeled Industrial Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing cross-domain transfer methods for industrial data prediction face challenges in regression tasks due to distribution biases and lack of labeled data in target domains, leading to decreased predictive performance.
Innovation Solution
A cross-domain transfer method involving pre-training a model in the source domain using contrastive domain generalization loss and self-supervised alignment, followed by adaptive processing in the target domain with instance-wise adversarial discrimination to align and calibrate the model, enhancing its predictive accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing cross-domain transfer methods are used for industrial data prediction, then the model can be transferred from source domain to target domain, but the predictive performance decreases due to distribution biases and lack of labeled data in target domain
Solution Approach 1:
The patent applies preliminary action by performing contrastive domain generalization loss calculation during the pre-training phase in the source domain. This preliminary alignment of latent features between domains reduces the distribution bias before the model is transferred to the target domain, thereby maintaining predictive performance despite domain differences
Solution Approach 2:
The patent introduces an intermediary mechanism through the contrastive domain generalization loss function, which acts as a mediator to align the latent feature distributions between source and target domains. This intermediary loss function bridges the distribution gap without requiring labeled data from the target domain
2Adaptability or versatility
If traditional transfer learning methods are applied to regression prediction tasks, then knowledge from source domain can be transferred, but most methods are designed for classification tasks and cannot be used for regression
Solution Approach 1:
The patent applies parameter changes by modifying the loss function parameters to suit regression tasks. The contrastive domain generalization loss is specifically designed with parameters (τ and ε) that control the temperature scaling and margin constraints, enabling the method to adapt from classification-oriented transfer learning to regression prediction tasks while maintaining prediction accuracy
3Productivity
If model pre-training is performed in source domain only, then training efficiency is improved, but the model fails to adapt to target domain data distribution
Solution Approach 1:
The patent performs preliminary domain alignment action during source domain pre-training by incorporating contrastive domain generalization loss. This preliminary alignment ensures that when the model is later applied to the target domain, it already has adapted features that are consistent with the target domain distribution, eliminating the need for extensive retraining
Solution Approach 2:
The patent implements feedback through the contrastive domain generalization loss calculation, which provides continuous feedback during pre-training about how well the latent features align with the target domain distribution. This feedback mechanism guides the model to adjust its features to better match the target domain, improving domain alignment accuracy while maintaining training efficiency
Data Source
AI summary
A cross-domain transfer method, apparatus and device for a prediction model, and a storage medium, including: acquiring source domain data in a source domain, determining a contrastive domain generalization loss according to a first latent feature and a label of the source domain data, pre-training a source domain model to obtain a pre-trained model, and transferring the pre-trained model to a target domain to make the pre-trained model adapted to the target domain and form a target domain model; acquiring target domain data, determining a second latent feature and a pseudo label of the target domain data, determining an instance-wise adversarial loss, a self-supervised alignment loss, and a pseudo domain generalization loss of the target domain data according to the second latent feature, pseudo label, source domain data, first latent feature and label, and performing calibrating processing on the target domain model to obtain a target model.