Federated Operating Model Grouping for Similar Technical Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning models often suffer from insufficient accuracy due to training solely on local data, which may not account for patterns from other devices, leading to low accuracy in anomaly detection and classification across heterogeneous datasets.
Innovation Solution
A method involving the generation of a basic device model, distribution to control devices, training, and aggregation using similarity functions to form group models, with a central server calculating centroid models and weighting group contributions for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train models locally on each device without data sharing, then data privacy and security are improved, but model accuracy deteriorates due to insufficient training data diversity
Solution Approach 1:
The patent merges models from multiple devices with similar characteristics into groups, combining their training results to create more accurate group models. This allows each device to benefit from aggregated knowledge of similar devices while maintaining data privacy, as only model parameters are shared within groups rather than raw data.
Solution Approach 2:
The patent changes the approach from sharing raw data to sharing model parameters (weights and biases). By exchanging parameter information instead of data, the system maintains data privacy while still enabling collaborative learning to improve model accuracy through parameter aggregation and averaging.
2Ease of operation
If traditional federated learning aggregates all device models equally, then implementation simplicity is improved, but model accuracy deteriorates due to heterogeneous data distributions across devices
Solution Approach 1:
The patent applies local quality by treating different device groups differently based on their similarity characteristics. Instead of uniform aggregation, devices are grouped by similarity metrics and aggregated separately, allowing each group to have tailored model aggregation that accounts for its specific data distribution characteristics.
Solution Approach 2:
The patent segments the heterogeneous device population into homogeneous groups based on similarity metrics. This segmentation allows the system to handle data heterogeneity by processing similar devices together, improving anomaly detection accuracy for each segment while managing complexity through modular group-based aggregation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Method for operating a technical system (A1-A3) with an optimal model, wherein the system (A1-A3) is part of a system (S) with a first technical system (A1) and at least one second technical system (A2, A3), each system (A1-A3) comprising a control device (15-17) and an associated technical device (25-27), and the system further comprising a server (FLS1) with a memory (MEM), wherein the following method steps are performed: Q1) generating a device base model (MB), Q2) distributing the device base model (MB) to the control devices (15-17), Q3) generating, training, and storing a first device model (MS1) by the control devices (15-17), Q4) providing and storing the device models (MS1-MS3) in a model memory (MEM) of the server (FLS1), Q5) loading and providing at least those device models (MS2, MS3) to the corresponding control device (15), which are not already present in the control device (15),Q6) Applying the device models (MS2, MS3) provided in step Q5) to the respective control device (15) and determining the respective similarity functions (SF1_2, SF1_3), Q7) Transmitting the respective similarity functions (SF1_2, SF1_3, SF2_1, SF2_3, SF_3_1, SF3_2) to the server (FLS1), Q8) Forming a group and creating a group model (MG1-MG3) through federated learning within the group using the similarity functions, Q9) Selecting and loading an operating model (M01) from the group models (MG1-MG3) for a plant (A1) and providing the operating model (M01) to its control device (15), as well as controlling the device (25) using the selected operating model (M01) as the optimal model.