Machine Controller Configuration Using Physics-Guided Residual Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine controllers require significant computational resources for real-time simulations, often resulting in higher simulation errors with reduced order models or the black-box nature of machine learning models leading to inaccurate predictions, especially in poorly covered machine states.

Innovation Solution

A machine controller configured with a machine learning module and simulation module that uses geometry and physical data to generate physical property values on a discretized grid, evaluates compatibility, and trains to minimize residuals, allowing for predictive control with reduced computational effort and increased accuracy by accounting for physical interrelationships.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If reduced order models are used for real-time simulations, then computational resources are reduced, but simulation errors increase

Engineering Contradiction:
Improvecomputational resourcesVSAvoidsimulation errors
Core Design Contradiction:
Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent introduces a physics-informed neural network as an intermediary between traditional simulation models and machine learning approaches. This hybrid model incorporates physical laws as constraints while using neural networks for efficient computation, thereby reducing computational resources while maintaining simulation accuracy through physics-based guidance.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the parameters of the simulation model by incorporating physics-informed constraints and loss functions that encode physical laws. This allows the model to operate with reduced computational complexity while maintaining accuracy by guiding the learning process with physical principles rather than relying solely on data-driven approaches.

Inventive Principle:
Principle #35Parameter changes

2Speed

If surrogate models based on machine learning are used, then computational speed is improved, but prediction accuracy deteriorates in poorly covered states

Engineering Contradiction:
Improvecomputational speedVSAvoidprediction accuracy
Core Design Contradiction:
SpeedVSReliability

Solution Approach 1:

The patent introduces physics-informed constraints as an intermediary mechanism that guides the surrogate model's predictions. By embedding physical laws into the neural network's loss function and architecture, the model remains physically consistent even when operating in states poorly covered by training data, thereby improving reliability while maintaining computational speed.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements feedback mechanisms where the physics-informed loss function continuously guides the surrogate model's predictions during inference. The model receives feedback from physical constraints that correct deviations from physically plausible behavior, ensuring accurate predictions even in extrapolation scenarios where training data is scarce.

Inventive Principle:
Principle #23Feedback

3Reliability

If physics informed neural networks are used, then physical constraints are satisfied, but re-training cost increases when physical constraints change

Engineering Contradiction:
Improvephysical constraints satisfactionVSAvoidre-training cost
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The patent implements dynamic physics-informed neural networks where physical constraints are formulated in a modular and adaptable manner. When physical constraints change, only the relevant portions of the loss function need to be updated rather than retraining the entire model from scratch, significantly reducing re-training costs while maintaining satisfaction of the new physical constraints.

Inventive Principle:
Principle #15Dynamics

4Measurement precision

If full order simulation models are used, then simulation accuracy is improved, but computational effort increases

Engineering Contradiction:
Improvesimulation accuracyVSAvoidcomputational effort
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts and separates the essential physical constraints from complex full-order simulation models and embeds them into a streamlined physics-informed neural network architecture. This extraction process captures the core physics in a computationally efficient form, achieving high simulation accuracy with significantly reduced computational effort compared to running full-order models.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3945375B1Machine controller and methods for configuring and using the machine controller
Publication Date: 2022.09.21 SIEMENS AG
  • EP3945375B1 patent drawingFigure 1~2

AI summary

For configuring a machine controller (CTL), geometry data (GD) and measured physical data (PD) of a machine (M) are received. The geometry data (GD) and the physical data (PD) are input to a machine learning module (NN) and to a simulation module (SIM) of the machine controller (CTL). By means of the input data, the simulation module (SIM) generates first values (K, F) of a first physical property of a component (CP) of the machine (M) on a discretized grid (GR). Furthermore, an evaluator (EV) is provided for evaluating a physical compatibility of the first values (K, F) with second values of a second physical property of the component (CP), and for generating a residual (D) quantifying the compatibility. The evaluator (EV) evaluates the compatibility of the first values (K, F) with output data (U) of the machine learning module (NN) and generates a resulting residual (D). Moreover, the machine learning module (NN) is trained to minimize the resulting residual (D), thus configuring the machine controller (CTL) for controlling the machine (M) by means of the output data (U) of the trained machine learning module (NN).