Control Model Execution Modules for Flexible Technical System Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity and variability of larger technical systems, such as gas turbines and production plants, pose challenges for the flexible use of learning-based control models due to difficulties in understanding internal effects and certification requirements, making their implementation cumbersome.
Innovation Solution
A method and device that utilize a data container with encoded control models and model-type-specific execution modules, allowing for the selection of appropriate execution modules and data channels based on model type information, enabling flexible implementation and operation of different control models across various systems, while ensuring data security through encryption and digital signatures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning control models are used to optimize complex technical systems, then system optimization capability is improved, but control system complexity and certification difficulty increase
Solution Approach 1:
The control system is segmented into multiple execution modules, each specialized for a specific machine learning model type (e.g., neural networks, support vector machines, decision trees). This segmentation allows each module to handle only its designated model type, simplifying the overall system architecture by dividing the complex task of handling diverse ML models into manageable, specialized components.
Solution Approach 2:
An intermediary layer is introduced between the diverse machine learning models and the technical system control interface. This intermediary consists of standardized execution modules that translate various ML model outputs into uniform control commands, thereby decoupling the complexity of different ML model implementations from the control system's core logic and certification requirements.
2Adaptability or versatility
If multiple different control models are implemented to handle varying requirements, then adaptability is improved, but implementation complexity increases
Solution Approach 1:
The execution modules are designed with universal interfaces that can handle multiple ML model types through a standardized framework. Each execution module is configured to work with specific model types, but the overall system provides universal adaptability by selecting and deploying the appropriate module based on the required model type, thus achieving versatility without proportionally increasing implementation complexity.
Solution Approach 2:
The system manages complexity by changing parameters such as model type identifiers, execution module selections, and configuration settings rather than fundamentally altering the control architecture. This allows the system to adapt to different ML model requirements by adjusting these parameters while maintaining a consistent underlying framework, thereby improving adaptability without linearly increasing implementation complexity.
3Adaptability or versatility
If learning-based control models are used, then control flexibility is improved, but understanding internal interactions becomes difficult
Solution Approach 1:
The standardized execution modules serve as intermediaries that provide a transparent interface between the black-box machine learning models and the control system. These modules maintain logs, standardized input-output interfaces, and configuration metadata that make the interactions between ML models and technical systems more observable and understandable, thereby reducing the difficulty of detecting and measuring internal interactions while preserving control flexibility.
4Productivity
If machine learning control models are deployed in larger technical systems, then system performance is improved, but certification process complexity increases
Solution Approach 1:
By segmenting the control system into standardized execution modules with defined interfaces and behaviors, the certification process can focus on verifying each module independently rather than the entire complex ML-based control system at once. This modular segmentation makes it easier to obtain certifications for individual components, thereby reducing overall certification complexity while maintaining high system performance.
Data Source
Figure 1~2
Figure 3~4
AI summary
In order to control a technical system (TS) by means of control model (SM, SM1, SM2) a data container (DC, DC1, DC2) is received, in which data container (DC, DC1, DC2) a control model (SM, SM1, SM2) having a training structure (TSR) and model type information (MTI, MTI1, MTI2) are encoded over all the model types. According to the invention, one of multiple model-type-specific execution modules (EM1, EM2, EM3) is selected for the technical system (TS) as a function of the model type information (MTI, MTI1, MTI2). Furthermore, operating data channels (BDC) of the technical system (TS) are assigned input channels (IC) of the control model (SM, SM1, SM2) as a function of the model type information (MTI, MTI1, MTI2). Operating data (BD) of the technical system (TS) are acquired via a respective operating data channel (BDC) and are transferred to the control model (SM, SM1, SM2) via an input channel (IC) assigned to this operating data channel (BDC). The control model (SM, SM1, SM2) is executed by means of the selected execution module (EM1, EM2), wherein control data (CD) are derived from the transferred operating data (BD) according to the training structure (TSR) and are output to control the technical system (TS).