Recurrent Neural Network Bottleneck for Technical System Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for controlling complex technical systems are not universally applicable and often fail to provide satisfactory results, as they rely on expert knowledge or reinforcement learning techniques that do not account for the dynamic behavior of systems effectively.

Innovation Solution

A method using a recurrent neural network coupled with a further neural network to model and predict the dynamic behavior of technical systems, allowing for the learning of an action selection rule that optimally controls the system by considering past and future states, with a bottleneck structure to focus on essential dynamics and account for manipulated variables.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If expert knowledge or reinforcement learning methods are used to control technical systems, then control decisions can be made automatically, but the methods are not universally applicable and do not provide sufficiently good results for complex systems with unpredictable dynamic behavior

Engineering Contradiction:
Improveautomatic controlVSAvoiduniversal applicability
Core Design Contradiction:
Extent of automationVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by creating a control method based on recurrent neural networks that can be applied to any technical system with dynamic behavior, regardless of the specific system type. The RNN-based predictor is designed to learn temporal patterns from training data and adapt to different systems, making the control approach universally applicable while maintaining automated decision-making capabilities

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If complex recurrent neural networks with multiple hidden layers are used to model dynamic behavior, then prediction accuracy improves, but computational complexity and data requirements increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidnetwork complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the control task into two distinct phases: offline training phase where the RNN is trained on historical training data to learn system dynamics, and online control phase where the trained network makes predictions. This segmentation allows the complex modeling work to be done beforehand, reducing real-time computational complexity while maintaining high prediction accuracy during actual control operations

Inventive Principle:
Principle #1Segmentation

3Reliability

If more training data is used to train the recurrent neural network, then the dynamic behavior modeling becomes more accurate, but the data processing time and computational resources increase

Engineering Contradiction:
Improvemodeling accuracyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by training the recurrent neural network offline using available training data before actual control operations begin. The RNN learns the system's dynamic behavior patterns during this offline phase, storing the knowledge in its weights and parameters. During online control, the pre-trained network makes rapid predictions without requiring additional training, thus achieving high reliability without time loss during critical control moments

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2112568B1Method for computer-supported control and/or regulation of a technical system
Publication Date: 2013.05.01 SIEMENS AG
  • EP2112568B1 patent drawingFigure 1
  • EP2112568B1 patent drawingFigure 2
  • EP2112568B1 patent drawing

AI summary

The invention relates to a method for the computer-aided control and/or regulation of a technical system. The method is characterized by two steps: learning the dynamics of a technical system using historical data based on a recurrent neural network, and subsequently learning an optimal control pattern by coupling the recurrent neural network with another neural network. The method according to the invention uses a recurrent neural network with a special hidden layer, which at any given time comprises a first hidden state and a second hidden state. The first hidden state is coupled to a matrix to be learned containing the second hidden state. In this way, a bottleneck structure can be created by choosing the dimension of the first hidden state to be smaller than the dimension of the second hidden state, or vice versa.This allows the network to better consider the essential autonomous dynamics of the technical system during learning and improves the network's approximation capability. The invention has a wide range of technical applications and can be used, in particular, for any technical system to optimally control these systems using computer-aided methods. One application area is, for example, the control of a gas turbine.