Text-Guided Sensor Data Prediction Across New Physical Quantities

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing models trained on specific physical quantities face challenges in adapting to new modeling tasks without significant parameter adjustments due to non-generalizable dimensions and dynamics.

Innovation Solution

A computer-implemented method utilizing a transformer-based model with text encoders and neural networks to adapt sensor data processing, enabling zero-shot or few-shot learning for new tasks by encoding sensor data characteristics and position, allowing for accurate prediction of sensor data without extensive retraining.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a model is trained on specific physical quantities, then the model achieves high accuracy for those specific quantities, but the model cannot generalize to new physical quantities without significant parameter adjustments

Engineering Contradiction:
Improveprediction accuracyVSAvoidgeneralizability to new physical quantities
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by designing a transformer model with text encoders that can process multiple types of sensor data across different physical quantities using the same architecture. The model learns generalizable representations of sensor data characteristics and temporal dynamics that transfer across different measurement types, enabling the single model to serve multiple functions for various physical quantities without requiring separate specialized models for each.

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

Solution Approach 2:

The patent employs parameter changes by using text encoders to dynamically represent different sensor data types and dimensions through textual descriptions. Instead of fixing model parameters for specific physical quantities, the system changes the textual parameters that describe the sensor data characteristics, allowing the model to adapt to new physical quantities by simply providing appropriate text descriptions rather than retraining with adjusted numerical parameters.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If many parameters are adjusted by fine-tuning to adapt to new modeling tasks, then the model can be adapted to new physical quantities, but the adaptation process becomes complex and time-consuming

Engineering Contradiction:
Improveadaptability to new tasksVSAvoidparameter adjustment complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent introduces text encoders as an intermediary layer between the input sensor data and the transformer model. These text encoders translate sensor data characteristics into textual representations that the model can process. This intermediary mechanism simplifies adaptation to new tasks because users only need to provide text descriptions of the new sensor data types rather than manually adjusting numerous model parameters, thereby reducing adaptation complexity while maintaining versatility.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If extensive retraining is performed to adapt models to new tasks, then the model achieves accuracy for new physical quantities, but the training time and computational resources increase significantly

Engineering Contradiction:
Improveprediction accuracy for new tasksVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the transformer model on a diverse set of sensor data types and physical quantities during an initial training phase. This pre-training establishes generalizable representations of temporal dynamics and sensor characteristics. When adapting to new tasks, the model leverages this preliminary knowledge and requires only minimal fine-tuning or inference with text descriptions, rather than performing extensive retraining, thereby significantly reducing the time and computational resources needed for adaptation.

Inventive Principle:
Principle #10Preliminary action

4Reliability

If the model uses fixed dimensions and learned dynamics for specific sensor data, then the model performs well for those data types, but the dimensions and dynamics do not generalize to new sensor data types

Engineering Contradiction:
Improvemodel performance consistencyVSAvoidgeneralizability of dimensions and dynamics
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamics by making the model's dimensional representations and learned dynamics adaptive rather than fixed. The text encoders dynamically generate representations based on the specific sensor data type being processed, allowing the model to adjust its internal dimensions and temporal dynamics to match the characteristics of each new sensor data type. This dynamic adaptation mechanism maintains reliable performance across different data types while enabling generalization to new sensor measurements.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4693104A1Device and computer-implemented method for processing sensor data
Publication Date: 2026.02.11 ROBERT BOSCH GMBH
  • EP4693104A1 patent drawingFigure 1~2
  • EP4693104A1 patent drawingFigure 3~4
  • EP4693104A1 patent drawingFigure 5

AI summary

Device and computer-implemented method for processing sensor data, in particular time series of sensor data, wherein the sensor data are provided in a first channel, the first channel comprising a first part of the sensor data, in particular a first time series of sensor data, wherein a first text (204) is assigned to the first channel, which characterizes the sensor data and/or a dimension of the sensor data in the first channel, wherein a first text encoding (404) is determined depending on the first text (204), in particular with a text encoder (402), wherein a first channel position encoding (408) is determined depending on the first text encoding (404), in particular with a neural network (406), and wherein a second text encoding (404) is determined depending on a given text, in particular with the or a text encoder (402).wherein the given text characterizes predictable sensor data and/or a dimension of predictable sensor data, wherein a second channel position coding (408) is determined depending on the second text coding (404), in particular with the or a neural network (406), wherein a first input variable (306) of an encoder is determined depending on sensor data from the first channel and depending on the first channel position coding (408), wherein a first input variable of a decoder (610, 710) is determined with the encoder depending on the first input variable (306) of the encoder (302), wherein a second input variable of the decoder is determined depending on the second channel position coding (408), and wherein sensor data, in particular a time series of sensor data, are predicted with the decoder depending on the first input variable of the decoder and the second input variable of the decoder.