Feedback Control with Limited Neural Output for Stable Alignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Controllers based on machine learning can output abnormal control commands in situations significantly different from the learning data, leading to unreliable control, particularly in large-scale deep neural networks, making it difficult to predict their response accurately.

Innovation Solution

A feedback control system incorporating a first control unit using PID control and a second control unit with a learning control unit, where a neural network adjusts parameters, and an adder combines their outputs, with an output limiter to constrain the second control unit's manipulated variable, ensuring stability and reliability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a controller based on machine learning is used to improve control performance, then control accuracy is improved, but reliability deteriorates because the controller may output abnormal commands in situations different from learning data

Engineering Contradiction:
Improvecontrol accuracyVSAvoidcontrol reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The control device is divided into multiple independent control units (first control unit, second control unit, third control unit), each responsible for different aspects of control. This segmentation allows the system to distribute the control function across multiple modules, reducing the risk that a single machine learning controller will cause system-wide failures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

A limiter is introduced as an intermediary component between the second control unit (machine learning-based) and the controlled object. The limiter acts as a safety mechanism that intercepts and restricts abnormal control commands before they can affect the controlled object, thus protecting system reliability while preserving the accuracy benefits of machine learning control.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If a large-scale deep neural network is used to improve control capability, then adaptability is improved, but reliability deteriorates because it becomes difficult to predict the controller's response

Engineering Contradiction:
Improvecontrol adaptabilityVSAvoidpredictability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The control system is segmented into multiple control units with different functions. The first control unit handles basic control tasks with predictable behavior, while the second control unit (deep neural network) handles complex adaptive tasks. This segmentation isolates the unpredictability of the deep neural network to a specific module, allowing the overall system to maintain reliability through the predictable behavior of other units and the limiter.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The limiter serves as an intermediary that mediates between the unpredictable deep neural network controller and the controlled object. It monitors and restricts the output of the neural network, ensuring that even when the neural network produces unpredictable responses, the controlled object receives only safe and appropriate commands.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If multiple control units are combined to improve reliability, then control reliability is improved, but device complexity increases

Engineering Contradiction:
Improvecontrol reliabilityVSAvoidcontrol system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The control device is divided into distinct functional modules (first control unit, second control unit, limiter, adder), each with a specific responsibility. This modular segmentation makes the complex system easier to understand, design, and maintain compared to a monolithic control system, as each module can be developed and tested independently.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Multiple control units are merged through an adder that combines their outputs. This merging approach allows the system to leverage the strengths of different control strategies (classic control and machine learning control) while managing complexity through a unified output combination mechanism.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3693798B1Processing apparatus and article manufacturing method
Publication Date: 2025.08.27 CANON KK
  • EP3693798B1 patent drawingFigure 1
  • EP3693798B1 patent drawingFigure 2
  • EP3693798B1 patent drawingFigure 3

AI summary

A feedback control device that takes information regarding a control deviation between a measured value and a desired value of a controlled object as input, and outputs a manipulated variable for the controlled object, includes: a first control unit that takes information regarding the control deviation as input, and outputs a manipulated variable for the controlled object; a second control unit that takes information regarding the control deviation as input, and that includes a learning control unit in which a parameter for outputting a manipulated variable for the controlled object is determined by machine learning; and an adder that adds a first manipulated variable output from the first control unit and a second manipulated variable output from the second control unit. A manipulated variable from the adder is output to the controlled object, and the second control unit includes a limiter that limits the second manipulated variable.