Sensor Data Transformer With Text Embeddings for Cross-Quantity Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models trained on specific physical quantities require extensive parameter adjustments for adaptation to new modeling tasks with different physical quantities, lacking generalization.
Innovation Solution
A computer-implemented method utilizing a transformer-based model with text encoders and neural networks to process sensor data, enabling adaptation to new tasks without extensive retraining by assigning text codes to sensor data dimensions and using channel and time position codings to predict sensor data accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a model trained on specific physical quantities is adapted to new modeling tasks with different physical quantities, then the model can handle new tasks, but extensive parameter adjustments and fine-tuning are required
Solution Approach 1:
The patent applies universality by designing a transformer model architecture that can process multiple types of physical quantities through unified text encoding. The model uses text embeddings to represent different physical quantity types, allowing the same model structure to handle diverse modeling tasks without requiring separate specialized models for each physical quantity type.
Solution Approach 2:
The patent changes parameters from fixed, task-specific weight matrices to learnable text embedding vectors. By replacing traditional parameter representations with text-based embeddings that capture semantic information about physical quantities, the model can adapt to new tasks by learning new embeddings rather than adjusting extensive parameters through fine-tuning.
2Reliability
If traditional models are used for multivariate time series forecasting, then they can process sensor data, but they lack generalization to new physical quantities
Solution Approach 1:
The patent introduces text embeddings as an intermediary layer between the input sensor data and the prediction mechanism. These text embeddings serve as semantic mediators that encode information about physical quantity types, enabling the model to generalize to new physical quantities by leveraging the semantic relationships captured in the text embeddings rather than relying on task-specific parameter adjustments.
Solution Approach 2:
The patent adds a semantic dimension to the traditional time series processing by incorporating text embeddings that represent physical quantity types. This transforms the problem from purely numerical time series analysis to a multi-dimensional approach that includes semantic information, allowing the model to generalize across different physical quantities by utilizing this additional semantic dimension.
3Measurement precision
If extensive fine-tuning is performed to adapt models to new tasks, then model performance improves, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary action by pre-training the transformer model on a diverse dataset encompassing multiple physical quantity types before deployment. This pre-training establishes a robust foundation of learned patterns and relationships that can be directly applied to new tasks with minimal additional training, eliminating the need for extensive fine-tuning when adapting to new physical quantities.
Solution Approach 2:
The patent enables self-service by designing a model that can automatically adapt to new physical quantities through text embedding learning without requiring manual parameter adjustment or extensive retraining. The model learns the semantic characteristics of new physical quantities directly from the text descriptions, performing self-adaptation with minimal human intervention and computational resources.
Data Source
AI summary
An apparatus and a computer-implemented method for processing time series of sensor data. The sensor data are provided in a first channel which includes a first time series of sensor data. A first text is assigned to the first channel which characterizes the sensor data and/or a dimension of the sensor data in the first channel. A first text coding is determined depending on the first text using a text encoder. A first channel position coding is determined depending on the first text coding using a neural network. A second text coding is determined depending on a predetermined text using a text encoder. The predetermined text characterizes sensor data to be predicted and/or a dimension of sensor data to be predicted. A second channel position coding is determined depending on the second text coding using a neural network.