Assembly Quality Prediction Using Privacy-Preserving Unified Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In manufacturing scenarios where multiple manufacturers participate in producing a single product, existing AI models struggle to predict assembly quality due to high data confidentiality requirements, decentralized and non-uniform data, and scarcity of data features.
Innovation Solution
A method and system for predicting product assembly quality using longitudinal unified learning, which involves sample alignment under an encryption policy, training local models, performing gradient security aggregation using homomorphic encryption, and merging data features to train a global model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If decentralized learning is used to create a global model by learning from multiple disperse edge clients, then data confidentiality is protected, but feature data becomes inconsistent across different production environments and detection devices
Solution Approach 1:
The patent applies parameter changes by transforming feature data through standardized processing parameters. Different feature data from various production environments and detection devices are converted into a unified parameter space, allowing consistent comparison and aggregation while preserving data confidentiality through decentralized learning.
Solution Approach 2:
The patent introduces an intermediary mechanism in the form of a centralized server that coordinates feature data aggregation without directly accessing raw confidential data. The server acts as a mediator that receives processed feature data from multiple edge clients, performs unified learning, and distributes the global model back to participants.
2Measurement precision
If data from multiple manufacturers is centrally processed to train an AI model, then model training quality is improved, but manufacturers are unwilling to disclose data information due to high data confidentiality requirements
Solution Approach 1:
The patent segments the centralized training process into decentralized local training operations at each manufacturer's edge device. Each participant trains local models using their own confidential data without sharing it, and only model parameters or gradients are aggregated centrally, thus maintaining both training quality and data confidentiality.
Solution Approach 2:
The patent uses copying by creating and aggregating copies of model parameters or gradients rather than original confidential data. The centralized server collects copies of learned features from multiple manufacturers, combines them to form a global model, without requiring access to the actual proprietary data from any single manufacturer.
3Adaptability or versatility
If sample data features are different across different product assembly or detection environments but sample IDs are the same, then data alignment becomes challenging, but unified learning requires consistent feature representation
Solution Approach 1:
The patent transforms diverse feature data from different environments into a unified parameter representation. By applying standardized parameter transformations and normalization, the system aligns features across different detection devices and production environments while maintaining the identity linkage through sample IDs.
Data Source
AI summary
In a method for predicting product assembly quality based on longitudinal unified learning, sample alignment is performed on a data sample of each participant to resolve problems of decentralization, nonuniformity, and scarcity of data; data partitioning is performed by a multi-player parallel structure on product assembly data by using a customized data partitioning policy, layer normalization is firstly performed by an encoder on partitioned data, and feature extraction is performed by a multi-thread attention layer to mine a correlation between each assembly production line inside the participant and assembly data of each device, so that a model feature extraction capability is enhanced; gradient security aggregation is performed on the local model of each participant by using a homomorphic encipherment method of secure multi-player computation, to obtain a global model, so that data of a plurality of sub-factories is merged to co-train a high-precision assembly quality prediction model.