Edge Gesture Prediction with Local Retraining and XAI

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques for determining gesture class predictions based on measurement data from sensors like radar sensors face challenges in accuracy and reliability due to mismatches between training scenarios and real-world deployment environments.

Innovation Solution

The implementation of edge-deployed machine learning models that enable local re-training using locally acquired training datasets, along with the use of feature relevance vectors determined through explainable artificial intelligence (XAI) techniques, to improve prediction accuracy and maintain user data privacy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional pretrained machine learning models are used for gesture class prediction, then the system can perform inference using centralized training, but the prediction accuracy degrades due to mismatch between training scenarios and real-world deployment environments

Engineering Contradiction:
Improveprediction accuracyVSAvoidadaptability to deployment scenarios
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system segments the training process into two phases: centralized pre-training phase and distributed edge re-training phase. The ML model is first trained centrally with initial training data, then deployed to multiple edge devices where each device performs local re-training with its own local training data. This segmentation allows the model to adapt to specific deployment environments while maintaining the benefits of centralized training resources.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system implements local re-training at each edge device using locally acquired training data specific to that deployment environment. Each edge device maintains a local training dataset that reflects its specific operational context, and the ML model is re-trained locally to adapt to these local conditions. This local quality approach ensures the model performs accurately in each specific deployment scenario rather than relying on a generic centralized model.

Inventive Principle:
Principle #3Local quality

2Reliability

If local re-training is performed at edge-deployed processors using locally acquired training datasets, then the ML model adapts to specific deployment scenarios improving prediction accuracy, but the complexity of the system increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The edge-deployed processor performs self-service by automatically collecting local training data from its operational environment and conducting local re-training of the ML model without requiring external intervention. The system autonomously manages the entire re-training pipeline including data collection, model updating, and performance monitoring, which simplifies deployment while maintaining high prediction accuracy through local adaptation.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements a feedback mechanism where the edge device continuously monitors prediction performance and uses this feedback to guide the local re-training process. The feedback loop ensures that the model adapts to local conditions while maintaining overall system reliability, and the automated feedback-driven approach reduces the need for manual system management despite the increased complexity of distributed re-training.

Inventive Principle:
Principle #23Feedback

3Loss of information

If measurement data is collected and processed locally at edge devices, then user data privacy is maintained, but the ability to perform centralized model updates and consolidations is limited

Engineering Contradiction:
Improveuser data privacyVSAvoidcentralized model management
Core Design Contradiction:
Loss of informationVSExtent of automation

Solution Approach 1:

The system extracts only the essential model updates (weight changes, gradient information) from the edge devices and sends these extracted updates to the central server, rather than transmitting raw measurement data or complete model states. This extraction approach maintains user data privacy by keeping sensitive local data at the edge while still enabling centralized model management through the transmission of condensed update information.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system introduces an intermediary layer that facilitates communication between edge devices and the central server. This intermediary manages the exchange of model updates, coordinates federated learning processes, and consolidates learnings from multiple edge devices without requiring direct access to raw measurement data. The intermediary enables centralized model management while preserving data privacy through controlled information exchange.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4542341A1Gesture class prediction and explainable artificial intelligence
Publication Date: 2025.04.23 INFINEON TECHNOLOGIES AG
  • EP4542341A1 patent drawingFigure 1
  • EP4542341A1 patent drawingFigure 2
  • EP4542341A1 patent drawingFigure 3

AI summary

Techniques of inference of gesture class predictions at edge-deployed devices (131, 132, 133, 134) are disclosed. The gesture class predictions are based on measurement data from radar sensors. A user output can be provided based on feature relevance vectors determined using explainable artificial intelligence. Edge-deployed re-training is disclosed. Techniques of federated learning at a server (139) are disclosed.