Ultrasonic Sensor Training Data Generation Using GAN-Based Synthesis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The generation of training data for ultrasonic sensor models is time-consuming and costly due to the need for numerous real-world environmental setups, especially when multiple objects are present, and existing machine-learning models require extensive hyperparameter tuning for generating high-quality synthetic data.

Innovation Solution

A method and device for generating synthetic training data using a generator model, such as a GAN or Variational Autoencoder, trained with scoring metrics to produce high-quality data for ultrasonic sensor models, reducing the need for real-world measurements and optimizing hyperparameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If real-world environmental setups are used for training data collection, then the quality and authenticity of training data is improved, but the time and cost required for data generation increases significantly

Engineering Contradiction:
Improvequality of training dataVSAvoidtime for data generation
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses a generator model (such as GAN or Variational Autoencoder) to create synthetic copies of real training data. The generator learns the distribution and characteristics of real ultrasonic sensor data and generates artificial training samples that mimic real-world scenarios, eliminating the need for extensive physical environmental setups while maintaining data quality for model training

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs preliminary training of the generator model using a small set of real data to establish the data distribution. Once trained, the generator can rapidly produce large volumes of synthetic training data without requiring repeated real-world measurements, thus reducing the time investment for subsequent data generation tasks

Inventive Principle:
Principle #10Preliminary action

2Reliability

If real-world environmental setups are used for training data collection, then the authenticity of training data is improved, but the cost of data generation increases

Engineering Contradiction:
Improveauthenticity of training dataVSAvoidcost of data generation
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The generator model creates synthetic replicas of authentic training data by learning from real sensor measurements. These synthetic copies capture the essential characteristics, noise patterns, and signal properties of real ultrasonic data, providing reliable training material without the high costs associated with setting up multiple physical test environments

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system varies parameters in the synthetic data generation process (such as object positions, distances, environmental conditions) to create diverse training scenarios. This allows comprehensive coverage of different operational conditions at minimal cost, as parameter variations in software are far cheaper than physical reconfiguration of test environments

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If extensive hyperparameter tuning is performed for the generator model, then the quality of synthetic data is improved, but the complexity and time of the training process increases

Engineering Contradiction:
Improvequality of synthetic dataVSAvoidcomplexity of training process
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent employs a discriminator model that provides feedback to the generator during training. The discriminator evaluates the authenticity of generated samples and guides the generator to improve its output quality. This adversarial feedback mechanism automates the hyperparameter optimization process, reducing manual tuning complexity while maintaining high synthetic data quality

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The training process dynamically adjusts hyperparameters and model weights through the adversarial training loop. The generator and discriminator models continuously adapt to each other, with hyperparameters being optimized automatically through the training dynamics rather than manual configuration, simplifying the overall process complexity

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4141807B1Method and device for generating synthetic training data for an ultrasonic sensor model
Publication Date: 2026.04.15 ROBERT BOSCH GMBH
  • EP4141807B1 patent drawingFigure 1a~1b
  • EP4141807B1 patent drawingFigure 2~3
  • EP4141807B1 patent drawingFigure 4

AI summary

The invention relates to a computer-implemented method for generating synthetic training data for training of a data-driven ultrasonic sensor model (43) for a given configuration of an ultrasonic sensor system (2) having multiple ultrasonic sensor devices (21), wherein the training data includes input data representing time-series data of received ultrasonic signals and output data indicating object characteristics of environmental objects in a sensing range of the ultrasonic sensor system (2); comprising the steps of: - Providing (S2) real training data obtained by a measurement of the given configuration of an ultrasonic sensor system (2); - Training (S4) of a generator model by means of a training model using the real training data; - Using (S9) the generator model (51) to generate the synthetic training data by applying a random noise vector (z) as input.