Event-Driven APL Edge Inference Under Power Constraints

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Advanced Physical Layer (APL) edge devices in industrial systems face power constraints, limiting their ability to perform machine learning tasks due to energy restrictions, which hampers their capability to analyze and make inferences from sensor data within hazardous environments.

Innovation Solution

An event-driven framework is applied to APL-based edge devices to invoke and deploy machine learning models, allowing them to analyze input data, make inferences, and take actions within the power constraints, enabling the devices to detect normal and abnormal conditions and control industrial systems effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If machine learning tasks are performed at the edge device, then communication bandwidth consumption is reduced, but power consumption increases beyond APL constraints

Engineering Contradiction:
Improvecommunication bandwidth consumptionVSAvoidpower consumption
Core Design Contradiction:
Loss of informationVSUse of energy by moving object

Solution Approach 1:

The patent segments the machine learning workload between cloud-based training and edge-based inference. The ML model is trained in the cloud using abundant resources, then deployed to the APL edge device for lightweight inference tasks. This segmentation allows the edge device to perform classification and decision-making with minimal power consumption while avoiding the need to transmit raw sensor data to the cloud, thus reducing communication bandwidth consumption on both ends.

Inventive Principle:
Principle #1Segmentation

2Productivity

If APL edge devices perform data analysis and machine learning, then real-time decision-making capability is improved, but the device exceeds its power budget

Engineering Contradiction:
Improvereal-time decision-making capabilityVSAvoidpower budget
Core Design Contradiction:
ProductivityVSUse of energy by stationary object

Solution Approach 1:

The patent applies parameter changes by optimizing the ML model for the specific power and computational constraints of APL edge devices. The model is designed with reduced complexity, using fewer parameters and simpler architectures that can run on limited hardware. This allows the device to perform real-time inference and decision-making within its power budget, enabling local classification of sensor data and immediate control actions without exceeding energy limitations.

Inventive Principle:
Principle #35Parameter changes

3Use of energy by stationary object

If APL edge devices have minimal processing resources, then power consumption is reduced, but machine learning task performance deteriorates

Engineering Contradiction:
Improvepower consumptionVSAvoidmachine learning task performance
Core Design Contradiction:
Use of energy by stationary objectVSProductivity

Solution Approach 1:

The patent applies preliminary action by pre-training the ML model in the cloud using powerful computing resources before deployment to the edge device. The model is trained offline on historical sensor data to learn patterns and decision boundaries. Once deployed, the pre-trained model performs inference locally with minimal computational requirements, maintaining high accuracy while consuming little power. This preliminary training action transfers the heavy computational burden from the resource-constrained edge device to the cloud.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4354231A1Systems and methods for implementing machine learning in a local APL edge device with power constraints
Publication Date: 2024.04.17 SCHNEIDER ELECTRIC SYSTEMS USA INC
  • EP4354231A1 patent drawingFigure 1
  • EP4354231A1 patent drawingFigure 2
  • EP4354231A1 patent drawingFigure 3

AI summary

A method performed by an Advanced Physical Layer (APL)-based edge device with power constraints is provided. The method includes applying an event-driven framework that is compliant with power constraints of the APL-based edge device to receive input data; applying the event-driven framework to the input data to invoke a machine learning (ML) model that is trained to analyze the input data and make inferences about one or more aspects of an industrial system based on the input data, and applying the invoked machine learning model to analyze the input data and make an inference about the one or more aspects of the industrial system based on the input data. The input data is received by the APL-based edge device from one or more source field devices of the industrial system and/or the inference is used make a decision and cause an action to be applied to the industrial system.