Linked Machine Learning Models for Low-Compute Industrial Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods for industrial monitoring and control devices require significant computational resources, storage, and network bandwidth for updating classification models, leading to high costs and inefficiencies, as they typically necessitate complete retraining for online updates.
Innovation Solution
A method involving the training of multiple machine learning models over distinct time regions, with incremental updates and linkage models, allowing for reduced computational demands and improved accuracy by leveraging separate data sets and plausibility coefficients, enabling efficient control device operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complete retraining of classification models is performed for online updates, then model accuracy is maintained or improved, but computational power requirements, storage capacity, and network bandwidth increase significantly
Solution Approach 1:
The patent segments the model updating process into two distinct phases: offline training phase where complete datasets are processed to build initial models, and online inference phase where pre-trained models are deployed for real-time predictions. This segmentation allows comprehensive training without continuous computational overhead during operation, resolving the contradiction between maintaining accuracy and reducing computational power usage.
Solution Approach 2:
The patent implements preliminary action by pre-training classification models offline using complete datasets before deployment. The models are prepared in advance with all necessary learning, so that during online operation, only inference is required without needing to retrain. This preliminary training action eliminates the need for continuous computational resources during runtime while maintaining model accuracy.
2Measurement precision
If complete retraining of classification models is performed for online updates, then model accuracy is maintained or improved, but storage capacity and network bandwidth requirements increase significantly
Solution Approach 1:
The patent segments data management into offline training datasets stored in bulk, and online operational data stored minimally. Only essential operational data needed for inference is retained in the system, while comprehensive training data is processed offline and discarded after model training completes. This segmentation dramatically reduces storage capacity requirements during online operation while preserving model accuracy.
Solution Approach 2:
The patent applies discarding and recovering by discarding complete training datasets after offline model training is completed. The training data is recovered (used) for model building, then discarded to free up storage space. During online operation, only minimal operational data is retained, achieving high model accuracy without proportionally high storage requirements.
3Measurement precision
If complete retraining of classification models is performed for online updates, then model accuracy is maintained or improved, but network bandwidth consumption and costs increase
Solution Approach 1:
The patent segments model training and data transmission into offline and online phases. All heavy data transmission and model updates occur offline during scheduled maintenance windows, utilizing network bandwidth when not critical for operation. During online operation, pre-trained models run locally with minimal network communication, dramatically reducing network bandwidth consumption while maintaining model accuracy through periodic offline updates.
Solution Approach 2:
The patent implements preliminary action by completing all model training and data synchronization offline before online deployment. Models are pre-updated with latest training data and algorithms before being deployed to production systems. This preliminary preparation eliminates the need for continuous network communication during online operation, reducing network bandwidth loss to only essential operational exchanges.
Data Source
AI summary
A method and device for machine learning wherein, in order to create favorable method conditions, an at least first machine learning model is trained via an at least first data set, a second machine learning model is trained via a second data set, an at least first prediction data set is formed via the trained at least first machine learning model, a second prediction data set is formed via the trained second machine learning model, a linking machine learning model is trained at least via the first prediction data set and the second prediction data set, a third prediction data set is formed via the linking machine learning model, and controlled variables for controlling a control apparatus are formed at least via the third prediction data set, such that the demand for computing power is reduced and the prediction accuracy and control accuracy are increased.