Multi-Sensor Fusion With Graph Label Correction for Noisy Labels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multi-sensor fusion methods are unable to effectively handle missing and noisy labels, which are prevalent in practical application scenarios, especially in data-limited scenarios, and lack a unified framework for semi-supervised learning and label correction.

Innovation Solution

The Graph regularized AutoFuse (GAF) method employs joint optimization for sensor-specific learning, label propagation, and classification, using sensor-specific autoencoders to learn latent representations and a label consistent classifier, with hyperparameters controlling the tradeoff between graph structural information and known label information, to address the challenges of missing and noisy labels in multi-sensor fusion.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual labeling of training data is performed to improve classification accuracy, then label quality improves, but time consumption and cost increase significantly

Engineering Contradiction:
Improvelabel qualityVSAvoidtime consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by using autoencoders to pre-process and extract features from raw sensor data before classification. This preliminary feature extraction and data representation learning reduces the complexity of subsequent labeling tasks, allowing semi-supervised learning to achieve better results with fewer manually labeled samples.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary mechanism - the autoencoder-based feature representation and graph regularization framework - that bridges raw unlabeled data and final classification. This intermediary structure enables the system to leverage unlabeled data effectively, reducing dependence on manually labeled data while maintaining high classification accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If all training data is labeled to improve model performance, then classification accuracy improves, but cost and complexity increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidlabeling complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system applies partial action by labeling only a portion of the training data (semi-supervised approach) rather than all data. The graph regularization and autoencoder framework enables effective learning from this partial labeled set, achieving good classification accuracy without the excessive cost and complexity of complete data labeling.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the parameter of label availability from complete to partial by introducing graph regularization terms that exploit data manifold structure. This parameter change allows the system to adapt to semi-supervised scenarios where only partial labels are available, reducing labeling complexity while maintaining performance.

Inventive Principle:
Principle #35Parameter changes

3Ease of operation

If traditional multi-sensor fusion methods are used, then processing simplicity is maintained, but ability to handle noisy and missing labels deteriorates

Engineering Contradiction:
Improveprocessing simplicityVSAvoidrobustness to noisy labels
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent creates a composite framework that integrates autoencoders, graph regularization, and multi-sensor fusion. This composite structure combines multiple technical components to achieve robustness against noisy and missing labels while maintaining reasonable processing complexity, outperforming traditional single-method approaches.

Inventive Principle:
Principle #40Composite materials

Solution Approach 2:

The proposed framework serves multiple functions simultaneously: it performs feature extraction via autoencoders, handles label noise through graph regularization, and fuses multi-sensor data. This multi-functionality makes the system universally applicable to various semi-supervised multi-sensor scenarios with noisy labels, enhancing reliability without sacrificing too much operational simplicity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

4Ease of manufacture

If semi-supervised learning is applied to reduce labeling cost, then cost decreases, but handling noisy and missing labels becomes more difficult

Engineering Contradiction:
Improvelabeling costVSAvoidhandling noisy labels
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent converts the harmful effect of noisy labels into a beneficial feature by using graph regularization that exploits the manifold structure of data. The graph structure helps identify and correct noisy labels by leveraging relationships among samples, turning the challenge of noise into an opportunity for robust semi-supervised learning with reduced labeling cost.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The system implements feedback mechanisms where the graph regularization term continuously refines label predictions by comparing against the learned data manifold structure. This feedback loop enables the system to iteratively correct noisy labels and improve classification performance, making semi-supervised learning more reliable despite lower labeling costs.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4290414B1Method and system for multi-sensor fusion in the presence of missing and noisy labels
Publication Date: 2024.11.27 TATA CONSULTANCY SERVICES LTD
  • EP4290414B1 patent drawingFigure 1
  • EP4290414B1 patent drawingFigure 2
  • EP4290414B1 patent drawingFigure 3

AI summary

This disclosure relates to a method and system for multi-sensor fusion in the presence of missing and noisy labels. Prior methods for multi-sensor fusion do not estimate and correct labels for learning effective models in semi-supervised learning methods. Embodiments of the present disclosure provides a method for learning robust sensor-specific autoencoder based fusion model by utilizing a graph structure to perform label propagation and correction. In the disclosed Graph regularized AutoFuse (GAF) method latent representation for each sensor is learnt using the sensor-specific autoencoders. Further these latent representations are combined and fed to a classifier for multi-class classification. The disclosure presents a joint optimization formulation for multi-sensor fusion where label propagation and correction, sensor-specific learning and classification are executed together.