AI Model Deployment on PLCs via Computation Graph Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional engineering systems for distributed control systems, such as manufacturing plants, face inefficiencies and errors due to manual configuration and lack of automated verification and validation, leading to increased time, cost, and effort, especially during equipment changes or maintenance, and require labor-intensive deployment of AI models on Programmable Logic Controllers (PLCs) with limited computational resources.
Innovation Solution
A method and system for generating an AI model that determines an AI function within an engineering framework, converts it into a processing format, and deploys it on a processing module, using an extraction and extension module to extend the computation graph architecture, enabling automated configuration and continuous communication with the AI framework for efficient training and updates, utilizing a communication adapter with a semantic data model to simplify the integration and adaptation of AI models on PLCs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual configuration and engineering methods are used for distributed control systems, then experts can configure equipment with domain knowledge, but the process requires increased time, cost, and effort while lacking automated verification and validation
Solution Approach 1:
The system enables self-service through automated configuration where the engineering system automatically performs configuration tasks without requiring manual expert intervention. The automated verification and validation mechanisms allow the system to self-check and self-correct engineering configurations, reducing both time and potential errors.
Solution Approach 2:
Manual mechanical configuration processes are replaced with automated computational systems. The patent substitutes human expert manual configuration with automated engineering software that performs configuration, verification, and validation tasks computationally, significantly reducing time while maintaining or improving accuracy.
2Power
If AI models are trained using traditional frameworks on powerful hardware, then high computational power and memory are available for training, but deployment on PLCs with limited resources requires labor-intensive steps and multiple iterations
Solution Approach 1:
The patent extracts the trained AI model from the training environment and separates the training process from deployment. The model is extracted in a format suitable for deployment on resource-constrained PLCs, eliminating the need to replicate powerful training hardware at the deployment site and reducing deployment complexity.
Solution Approach 2:
An intermediary conversion process is introduced between training and deployment. The system uses a intermediate representation format that bridges the gap between powerful training hardware and constrained PLC deployment environments, simplifying the deployment process while maintaining model integrity.
3Adaptability or versatility
If conventional engineering systems rely on manual configuration, then flexibility in handling diverse equipment is maintained, but automated updates for new equipment during maintenance or downtime are not possible
Solution Approach 1:
The configuration system transitions from static manual configuration to dynamic automated configuration. The system automatically adapts to new equipment during maintenance or downtime by dynamically generating and updating configuration parameters based on equipment specifications and system requirements, maintaining flexibility while enabling automation.
Solution Approach 2:
The system performs preliminary automated configuration actions before equipment is fully integrated. Configuration templates and parameters are prepared in advance, allowing rapid automatic configuration when new equipment is inserted during maintenance or downtime, reducing overall integration time while maintaining adaptability.
4Measurement precision
If multiple iterations of labor-intensive deployment steps are executed for AI model updates, then model accuracy can be improved, but time and computational resources are significantly consumed
Solution Approach 1:
The patent uses copying by creating a simplified replica or representation of the complex training process that can be executed efficiently on PLCs. Instead of repeatedly executing full training iterations on resource-constrained devices, the system copies essential model updates and parameters, achieving accuracy improvements with significantly reduced computational overhead and time.
Data Source
AI summary
A method includes determining an artificial intelligence function in an engineering framework system. An inference path is defined for generation of an AI model by a computation graph. An AI function and the inference path are converted into a processing format. The converted AI function is sent and exported to an extraction and extension module of an AI workbench module. The extended computation graph of the inference path is transmitted from the extraction and extension module to an AI framework module. The method includes communicating of a communication adapter with the processing module continuously by using a supporting communication protocol for receiving training data as input for the AI function and forwarding the training data to the AI framework module. Learned parameters of the AI model are transferred from an API interface of the AI framework module to the communication adapter for updating the AI model.