Training Dataset Generation for Wall Diagnostic AI Using Pseudo-Labels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for generating training data sets for wall diagnostic devices using artificial intelligence are inadequate, leading to imprecise and unreliable classifications of wall types and objects within walls.

Innovation Solution

A method involving a classifier module trained via semi-supervised learning to classify unclassified sensor data, utilizing both classified sensor data and sensor data with pseudo-classifications, and employing latent space representations and distance determination to enhance classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional supervised learning is used to train the classifier module, then the training process requires large amounts of manually classified sensor data, but this increases the time and resources required for data preparation and limits the size of the training dataset

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata preparation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system enables unclassified sensor data to classify itself through the pseudo-classification module, which automatically generates pseudo-labels without human intervention. This self-service mechanism allows the training dataset to expand using unlabeled data, eliminating the time-consuming manual annotation process while maintaining reliable classification through the semi-supervised learning framework.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The pseudo-classification module performs preliminary classification of unclassified sensor data before the main training process. By pre-processing the unlabeled data and assigning pseudo-labels in advance, the system prepares a larger training dataset that can be used immediately, reducing the overall time required for data preparation while improving classification reliability.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If only manually classified sensor data is used for training, then the training dataset size is limited, but this reduces the statistical significance and generalization capability of the artificial intelligence

Engineering Contradiction:
Improvetraining dataset sizeVSAvoidclassification precision
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system merges manually classified sensor data with unclassified sensor data that has been assigned pseudo-labels by the pseudo-classification module. This combination creates a larger training dataset that maintains the quality of manually labeled data while incorporating the quantity of unlabeled data, thereby improving statistical significance and generalization capability without compromising classification precision through the semi-supervised learning approach.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The pseudo-classification module acts as an intermediary between unclassified sensor data and the training process. It generates pseudo-labels that bridge the gap between unlabeled and labeled data, enabling the system to utilize both types of data effectively. This intermediary mechanism allows the training dataset to expand while maintaining measurement precision through the carefully designed pseudo-labeling process.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If semi-supervised learning is implemented with pseudo-classifications, then the training dataset can be significantly enlarged, but this introduces potential errors from automated pseudo-labeling

Engineering Contradiction:
Improvetraining dataset generation efficiencyVSAvoidclassification reliability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system implements feedback mechanisms where the classifier module's predictions are continuously refined based on the pseudo-classified data. The semi-supervised learning process uses the pseudo-labeled data to improve the classifier, which in turn generates better pseudo-labels for subsequent iterations. This feedback loop enhances classification reliability while maintaining high productivity in training dataset generation.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system uses pseudo-classifications selectively rather than relying entirely on them. By combining a portion of manually classified data with pseudo-classified data, the system achieves sufficient training dataset size without fully depending on potentially error-prone automated labeling. This partial use of pseudo-classifications maintains reliability while improving productivity.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4618039A1Method for generating a training dataset
Publication Date: 2025.09.17 ROBERT BOSCH GMBH
  • EP4618039A1 patent drawingFigure 1
  • EP4618039A1 patent drawingFigure 2
  • EP4618039A1 patent drawingFigure 3

AI summary

The invention relates to a computer-implemented method (300) for generating a training data set (143) for training an artificial intelligence (125) for operating a measuring device (100), in particular a wall diagnostic device, comprising: receiving (301) unclassified sensor data (172) of at least one sensor unit (101) of a measuring device (100) by a classifier module (183); classifying (303) the unclassified sensor data (172) and providing classified sensor data (174) by a classifier module (183); and adding (305) the classified sensor data (174) to a training data set (143).