Federated Soft Sensor Modeling Without Gradient Instability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In complex industrial processes, the lack of data exchange among multi-party enterprises hinders the effectiveness of artificial intelligence models for predicting product quality, and federated learning faces challenges such as local minima, initial parameter dependence, and gradient issues, limiting the performance of neural networks.
Innovation Solution
A method utilizing a federated stochastic configuration network for training, where each factory initializes local models, uploads best candidate nodes to a central server for weighted aggregation and greedy selection, and distributes a trained global model as a soft sensor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train models from local data of multiple participants, then data privacy is protected, but the model training falls into local minima and suffers from gradient disappearance and gradient explosion
Solution Approach 1:
The patent replaces the traditional deep learning gradient-based optimization mechanism with a stochastic configuration mechanism. Instead of using backpropagation and gradient descent to train neural networks in federated learning, the invention employs random weight assignment and stochastic configuration of hidden layer nodes, eliminating gradient calculation entirely. This substitution resolves the gradient disappearance and explosion problems while maintaining data privacy protection through federated learning architecture.
Solution Approach 2:
The patent changes the training parameters and optimization approach by using random initialization and stochastic configuration instead of fixed gradient-based optimization. The method dynamically configures hidden layer parameters randomly and uses least squares estimation for output weights, avoiding the local minima problem inherent in gradient-based methods. This parameter change enables the model to achieve better generalization performance while protecting data privacy.
2Productivity
If deep learning algorithms are used for model training, then learning ability is strong, but the algorithms depend strongly on initial parameters and fall into local minima
Solution Approach 1:
The patent substitutes the gradient-based optimization mechanism with a stochastic configuration mechanism that uses random weight assignment and least squares estimation. This replacement eliminates the dependency on initial parameters and avoidance of local minima while maintaining strong learning ability through the universal approximation property of the stochastic configuration network.
Solution Approach 2:
The patent enables the network to automatically configure its hidden layer parameters through stochastic mechanisms without requiring manual initialization or tuning. The random configuration and least squares estimation allow the model to self-adapt to the data distribution, achieving reliable performance without strong dependence on initial parameter settings.
3Measurement precision
If more hidden layer nodes are added to improve model accuracy, then prediction performance improves, but model complexity and training time increase
Solution Approach 1:
The patent applies partial action by selectively configuring only the necessary hidden layer nodes through stochastic mechanisms rather than using excessive nodes. The random configuration process naturally prunes unnecessary connections and focuses on meaningful feature representations, achieving good prediction accuracy with simpler model structures and reduced training complexity.
Data Source
AI summary
Provided is an industrial process soft sensor method based on a federated stochastic configuration network, including: acquiring historical industrial process auxiliary data and corresponding product quality data; finding out optimal hidden layer parameters; processing, by a central server, to obtain global parameters, and downloading same to each factory as hidden layer parameters for a local model; obtaining output weights of a current network through an optimization algorithm, and uploading same to the server for weighted aggregation; and when the number of hidden layer nodes in a current network exceeds a maximum given value or a residual in current iteration meets an expected tolerance, completing modeling to obtain the global federated stochastic configuration network without adding new nodes. The present disclosure effectively improves prediction performance of models and protects data privacy, thus meeting industrial process soft sensor requirements.