Runtime Model Injection for Selective ML Controller Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The retraining of machine learning models is complex and time-consuming, leading to suboptimal and unreliable control of technical units, such as industrial robots and autonomous vehicles, due to domain drifts and the need for switching between high-performance and high-assurance controllers, resulting in safety hazards and operational inefficiencies.

Innovation Solution

A technical injection system comprising two computing units, where one unit preprocesses and stores a retrained machine learning model independently and injects the relevant parts into the second unit at runtime, allowing seamless and reliable updates without interrupting the control process, using an injection interface to replace the core of the current model with the retrained model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the machine learning model is retrained to adapt to domain drift, then the model performance and reliability are improved, but the retraining process is complex and time-consuming

Engineering Contradiction:
Improvemodel performanceVSAvoidretraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent segments the machine learning model into multiple independent components or layers, allowing selective retraining and injection of specific parts rather than retraining the entire model. This reduces the time and computational resources required while maintaining model performance.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary retraining of model components in advance, prepares the retrained parts, and stores them for later injection. This allows the retraining process to be completed beforehand, reducing the time required during actual deployment when domain drift is detected.

Inventive Principle:
Principle #10Preliminary action

2Stability of the object's composition

If the high-assurance controller is used longer to allow model retraining, then system stability is maintained, but the control reliability decreases due to suboptimal outputs

Engineering Contradiction:
Improvecontrol stabilityVSAvoidcontrol reliability
Core Design Contradiction:
Stability of the object's compositionVSReliability

Solution Approach 1:

The patent extracts only the necessary parts of the retrained model that need to be updated, rather than replacing the entire model. This allows for faster injection and reduces the time the system must rely on suboptimal controllers, thereby maintaining both stability and reliability.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces an intermediary injection mechanism that facilitates the transfer and integration of retrained model parts into the running system. This intermediary process enables seamless updates without requiring prolonged use of backup controllers, thus maintaining control reliability while ensuring stability.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If control switches between high-performance and high-assurance controllers, then model reliability is maintained, but energy consumption increases and operational stability decreases

Engineering Contradiction:
Improvemodel reliabilityVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent implements a dynamic model injection system that adapts to changing conditions by injecting retrained model parts only when necessary. This dynamic approach avoids unnecessary switching between controllers, reducing energy consumption while maintaining model reliability through targeted updates.

Inventive Principle:
Principle #15Dynamics

4Adaptability or versatility

If the entire machine learning model is replaced during retraining, then the model adapts to new data distributions, but the complexity and time required for retraining increases significantly

Engineering Contradiction:
Improvemodel adaptabilityVSAvoidretraining complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent divides the machine learning model into separable components that can be independently retrained and injected. This segmentation reduces retraining complexity by allowing focused updates on specific parts rather than the entire model, while still achieving adaptability to new data distributions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary retraining of model components and prepares them for injection in advance. This preliminary action reduces the complexity and time of the actual model update process, while ensuring the model adapts to new data distributions through pre-prepared retrained parts.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4220491A1Technical injection system for injecting a retrained machine learning model
Publication Date: 2023.08.02 SIEMENS AG
  • EP4220491A1 patent drawingFigure 1
  • EP4220491A1 patent drawingFigure 2
  • EP4220491A1 patent drawing

AI summary

The invention is directed to a technical injection system (1) for injecting a retrained machine learning model (30), comprising a. a first computing unit (10) comprising a first storage medium (12), wherein the first computing unit (10) is configured for providing the retrained machine learning model (30); and preprocessing the retrained machine learning model (30); wherein the retrained machine learning model (30) is stored in the first storage medium (12); b. a second computing unit (20) comprising a second storage medium (22) and an injection interface (40), wherein the injection interface (40) is configured for injecting at least one relevant part of the retrained machine learning model (30) after processing from the first storage medium (12) of the first computing unit (10) into the second storage medium (22) of the second computing unit (20) by means of the injection interface at runtime; wherein a current machine learning model is stored in the second storage medium (22); and the injection comprises the identification of the at least one relevant part of the current machine learning model and the replacement of the identified at least one relevant part of the current machine learning model by the corresponding at least one relevant part of the retrained machine learning model.