Interpretable Neural PID Controller for Validated Nonlinear Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The lack of interpretability in artificial neural networks used for control applications, particularly in high-stakes fields like high-power converter circuits, makes it difficult to formulate an adequate validation strategy, leading practitioners to prefer classical regulation schemes despite their limitations in handling nonlinearities.

Innovation Solution

The development of interpretable artificial neural networks that implement reinforcement learning strategies, specifically designed for nonlinear control applications, providing a mathematically sound validation approach by integrating proportional-integral-derivative (PID) or proportional-integral (PI) control functions within a neural network framework, allowing for the calculation of controller output signals based on error inputs and incorporating neural networks to estimate integral and differential components.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If classical regulation schemes like PID control are used, then reliability and interpretability are improved, but the ability to handle nonlinearities and adaptability deteriorates

Engineering Contradiction:
Improvevalidation strategyVSAvoidhandling nonlinearities
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent merges classical PID control structure with neural network learning capabilities. The controller combines proportional, integral, and derivative terms with neural network-based adaptive components, allowing the system to maintain the interpretability and reliability of classical control while gaining the adaptability to handle nonlinearities through reinforcement learning.

Inventive Principle:
Principle #5Merging (Combining)

2Adaptability or versatility

If reinforcement learning techniques are used, then adaptability and handling of nonlinearities are improved, but interpretability and reliability deteriorate

Engineering Contradiction:
ImproveflexibilityVSAvoidvalidation strategy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the control algorithm into distinct interpretable components: proportional term, integral term, derivative term, and neural network-based adaptive terms. Each component has a clear functional role and can be individually validated, maintaining overall system interpretability while enabling sophisticated nonlinear control through the neural network segments.

Inventive Principle:
Principle #1Segmentation

3Device complexity

If PID control is used for highly nonlinear systems, then simplicity is improved, but manufacturing precision and handling of nonlinearities deteriorate

Engineering Contradiction:
Improvecontrol schemeVSAvoidcontrol precision
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent transforms the static PID controller into a dynamic adaptive controller by incorporating reinforcement learning. The controller parameters are no longer fixed but adapt dynamically based on system state and performance feedback, enabling the simple PID structure to achieve high precision in nonlinear systems through learned parameter adjustments.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4167140A1Interpretable neural networks for nonlinear control
Publication Date: 2023.04.19 INFINEON TECH AUSTRIA AG
  • EP4167140A1 patent drawingFigure 1~2
  • EP4167140A1 patent drawingFigure 3~4
  • EP4167140A1 patent drawingFigure 5~6

AI summary

A controller circuit implements an interpretable neural-network-based proportional integral derivative (PID) control function. The controller circuit comprises a controller output signal for input to a nonlinear plant, a controller input signal representing an error in an output of the nonlinear plant, and a neural network configured to calculate the controller output signal from the controller input signal by summing a first signal depending on a current value of the controller input signal, a second signal generated at least in part by a first neural network estimating a differential of the controller input signal, and a third signal generated at least in part by a second neural network estimating an integral over time of the controller input signal.