AI Model Transfer Using Fourier-Based Device Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models for industrial devices are difficult to transfer between different environments due to variations in installation conditions and operating parameters, requiring extensive retraining for each new installation.
Innovation Solution
A method involving Fourier transforms of sensor signals from reference and target devices in different environments to generate a transfer function, allowing an AI model trained in one environment to be reused across identical devices without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an AI model is trained extensively for a specific device installation, then the model accuracy for that specific installation is improved, but the model cannot be transferred to other installations without extensive retraining
Solution Approach 1:
The patent introduces an intermediary representation layer between the raw sensor data and the AI model predictions. This intermediary involves transforming sensor signals through Fourier transforms and using domain adaptation techniques to create environment-invariant features. The transfer function acts as a mediator that maps features from source domain to target domain, enabling model transfer without retraining while preserving accuracy.
Solution Approach 2:
The patent changes the parameter space by transforming data from time domain to frequency domain using Fourier transforms. This parameter transformation allows the model to operate on invariant features that are independent of environmental variations. The transfer function further transforms the parameters by adapting feature distributions between different operating environments, enabling the model to maintain accuracy across installations.
2Reliability
If extensive retraining is performed for each new installation, then the model adapts to the new environment, but the time and computational resources required increase significantly
Solution Approach 1:
The patent performs preliminary actions by training the AI model extensively on a source installation and storing the trained model weights and the transfer function. This preliminary training captures the essential patterns that can be transferred to target installations. The transfer function is pre-computed based on domain adaptation theory, allowing rapid deployment to new installations without repeating the extensive training process.
Solution Approach 2:
The patent creates a copy of the trained model that can be deployed to multiple target installations. Instead of creating new models for each installation, the same trained model is copied and applied to different environments using the transfer function for adaptation. This copying approach maintains reliability across installations while eliminating the need for time-consuming retraining.
3Productivity
If the same AI model is used across different operating environments, then deployment speed is improved, but the model accuracy decreases due to environmental variations
Solution Approach 1:
The patent transforms the data parameters from time-domain sensor signals to frequency-domain representations using Fourier transforms. This parameter change creates environment-invariant features that maintain accuracy across different operating conditions. The frequency-domain representation captures essential device characteristics while filtering out environment-specific variations, enabling accurate predictions with rapid deployment.
Solution Approach 2:
The patent replaces the mechanical retraining process with a mathematical transformation approach. Instead of mechanically retraining the model on new data (which is time-consuming), the system uses Fourier transforms and transfer functions to mathematically adapt the model to new environments. This substitution maintains accuracy while enabling rapid deployment without iterative training processes.
Data Source
Figure 1
Figure 2~3g
AI summary
A computer-implemented method for operating a technical device (P2) using an artificial intelligence-based model, comprising the following steps: a) Acquiring a first operating parameter of a reference device (P1), which is identical in construction to the technical device (P2), using a first measuring instrument (MD1) in the form of a time-discrete first sensor signal in a first operating environment (E1) and a first permissible operating mode; b) Acquiring a second operating parameter of the reference device (P1) using a second measuring instrument (MD2) in the form of a time-discrete second sensor signal in a second operating environment (E2) and a second permissible operating mode; c) Performing a first Fourier transform (FFT1) for the first sensor signal and a second Fourier transform (FFT2) for the second sensor signal.d) Determining the quotient of the first Fourier transform and the second Fourier transform as a transfer function (F), e) Generating and training a first artificial intelligence-based model for the reference device in the first operating environment (E1), f) Generating a second artificial intelligence-based model for the technical device (P2) in the second operating environment (E2) using the first model and applying the transfer function (F), g) Operating the technical device (P1) with the second model.