Linked Machine Learning Models for Low-Compute Control Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models in industrial monitoring and control devices, such as those used in SCADA systems, often require complete retraining for updates, which is computationally expensive and bandwidth-intensive, making real-time updates impractical.

Innovation Solution

The method involves forming a second machine learning model based on the evaluation of the first model's accuracy and reliability, using non-overlapping data sets to reduce computational and bandwidth requirements, and combining these models into a linking model for enhanced forecasting and control accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If complete retraining of machine learning models is performed for updates, then model accuracy and reliability are maintained, but computing power requirements, storage capacity, and network bandwidth increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputing power requirement
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent segments the model update process into two distinct phases: an offline phase where the base model is trained comprehensively using historical data, and an online phase where only incremental updates are performed using new data. This segmentation allows the system to maintain model accuracy while significantly reducing the computing power and resources required during online operation, as the intensive training work is completed beforehand.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by training the base model offline before deployment. During the offline phase, the system performs comprehensive model training using extensive historical data, preparing the model in advance. When deployed online, the pre-trained model requires only minimal incremental updates, thereby reducing the computing power and bandwidth requirements during real-time operation while maintaining high accuracy.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If complete retraining of machine learning models is performed for updates, then model accuracy and reliability are maintained, but network bandwidth and transmission data increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidnetwork bandwidth consumption
Core Design Contradiction:
ReliabilityVSLoss of substance

Solution Approach 1:

The patent segments the data processing into offline training data (historical data used for base model training) and online update data (new data used for incremental updates). This segmentation reduces network bandwidth consumption online because only small increments of new data need to be transmitted and processed, rather than retransmitting the entire training dataset for complete retraining.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

By performing comprehensive model training offline using historical data before deployment, the system prepares the base model in advance. During online operation, only incremental updates with new data are required, significantly reducing the amount of data that needs to be transmitted over the network while maintaining model accuracy.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If machine learning models are updated online with active data transmission, then real-time accuracy is improved, but computing power and storage requirements increase

Engineering Contradiction:
Improveforecast accuracyVSAvoidcomputing power requirement
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the model into a pre-trained base model and incremental update components. The base model is trained offline with comprehensive historical data to capture general patterns, while online operations only require processing new data through the pre-trained model for incremental adjustments. This segmentation enables real-time forecast accuracy without requiring full retraining computing power during online operation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary comprehensive training of the base model offline before deployment. During online operation, the pre-trained model can immediately provide accurate forecasts with minimal additional computing requirements, as the heavy lifting of learning fundamental patterns has already been completed in the offline phase.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3662418B1Method and apparatus for machine learning in a computational unit
Publication Date: 2021.08.25 SIEMENS AG
  • EP3662418B1 patent drawingFigure 1
  • EP3662418B1 patent drawingFigure 2

AI summary

The invention relates to a method and to a device for machine learning. In order to create favorable method conditions, according to the invention, an at least first machine learning model (8) is trained by means of an at least first data set (3), a second machine learning model (9) is trained by means of a second data set (4), an at least first prediction data set (12) is formed by means of the trained, at least first machine learning model (8), a second prediction data set (13) is formed by means of the trained second machine learning model (9), a linking machine learning model (15) is trained at least by means of the first prediction data set (12) and the second prediction data set (13), a third prediction data set (14) is formed by means of the linking machine learning model (15), and controlled variables for controlling a control apparatus (2) are formed at least by means of the third prediction data set (14). Thus, the demand for computing power is reduced and the prediction accuracy and control accuracy are increased.