Physics-Informed Neural Networks for Converter Feedback Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional modeling techniques for switching-mode power converters struggle with accuracy due to parasitic effects and non-linearities, leading to overshoot, prolonged settling times, and oscillatory behavior, while data-driven machine learning models lack physical constraints and are prone to overfitting.
Innovation Solution
A physics-informed neural network (PINN) that integrates machine learning with physical principles to model feedback loop responses, using a ResNet-based encoder, physics-guided decoder, and output head to predict poles, zeros, and gain, ensuring physically meaningful outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional modeling techniques are used for switching-mode power converters, then the modeling process is simple, but accuracy deteriorates due to parasitic effects and non-linearities
Solution Approach 1:
The patent introduces an autoencoder neural network as an intermediary between raw converter data and the final model parameters. The autoencoder compresses input data through encoding layers to extract essential features, then reconstructs them through decoding layers to generate accurate pole-zero-gain representations, bridging the gap between simple data collection and accurate complex modeling
Solution Approach 2:
The patent replaces traditional mechanical/mathematical modeling approaches (circuit analysis, transfer function derivation) with a data-driven neural network system. Instead of manually deriving models from circuit equations, the system learns models directly from operational data, substituting conventional engineering methods with machine learning while maintaining physical interpretability through pole-zero-gain outputs
2Adaptability or versatility
If data-driven machine learning models are used, then adaptability improves, but physical constraints are violated leading to overfitting and non-physical predictions
Solution Approach 1:
The patent applies local quality by assigning specific physical meanings to different parts of the neural network output. The decoder layers are designed to output pole locations, zero locations, and gain values separately, each constrained to physically meaningful ranges. This localized physical interpretation ensures that each component of the model adheres to physical principles while maintaining overall system adaptability
Solution Approach 2:
The patent implements feedback through the autoencoder architecture where the decoded pole-zero-gain parameters are fed back to validate physical consistency. The loss function incorporates constraints that penalize non-physical predictions, creating a feedback mechanism that guides the network to produce reliable, physically consistent models while maintaining adaptability to different operating conditions
3Loss of time
If traditional modeling approaches are used, then computational resources are saved, but settling time increases due to overshoot and oscillatory behavior
Solution Approach 1:
The patent applies preliminary action by pre-training the autoencoder on comprehensive converter data covering various operating conditions. This pre-training establishes a robust foundation that enables fast, accurate predictions during deployment without requiring complex real-time computations. The model learns optimal pole-zero-gain configurations in advance, allowing rapid adaptation to new conditions with minimal computational overhead
Data Source
AI summary
Physics-informed neural networks and methods for modeling and analyzing feedback loop responses in electronic devices, including switching-mode power converters. The physics-informed neural network integrates physical principles with machine learning techniques to predict high-order transfer function parameters, such as poles, zeros, and gain, based on transient signals. These parameters are used to generate predicted frequency responses, including observable data like gain and phase, and non-observable data like poles and zeros. The predicted frequency responses are further translated into graphical representations, such as Bode plots and pole-zero plots, providing insights into system stability and performance. By extracting features that represent dynamic behavior and stability factors, the physics-informed neural network ensures predictions are physically meaningful and interpretable. The technology can be useful for real-time analysis, stability assessment, and automated compensation tuning in areas such as power management devices, industrial automation controllers, precision signal processing systems, and robotics platforms.