Generative Time-Series Modeling With Physical Machine Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining operating variables of machines struggle with accuracy and reliability, particularly when using probabilistic models, as it is challenging to ensure that predicted variables are physically viable.
Innovation Solution
A computer-implemented method and machine learning system that utilizes a generative model with multiple layers, including a first layer trained to map probabilistic noise to intermediate data, a second layer to map this data to a time series, and a third layer providing physical constraints learned from machine states and environment variables, thereby enhancing the accuracy and reliability of time series predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If probabilistic models are used to determine operating variables, then unobservable behavior can be modeled, but it is difficult to prove that the predicted variable is physically viable
Solution Approach 1:
The patent introduces an intermediary verification mechanism that acts as a mediator between the probabilistic model's predictions and physical reality. This intermediary system checks whether predicted operating variables satisfy physical constraints before accepting them as valid, thus resolving the contradiction between modeling flexibility and physical reliability.
Solution Approach 2:
The patent implements a feedback loop where predictions from the probabilistic model are continuously validated against physical constraints. When predictions violate physical viability, the system provides feedback to adjust the model or re-generate predictions, ensuring that only physically viable variables are output while maintaining the ability to model unobservable behavior.
2Reliability
If physical models are used to determine operating variables, then deterministic results are obtained, but accuracy depends on the observability of relevant information
Solution Approach 1:
The patent merges the strengths of both physical models and probabilistic models into a hybrid approach. The physical model provides deterministic constraints and structure, while the probabilistic model fills in unobservable variables and provides flexibility. This combination achieves both deterministic reliability and high measurement precision by leveraging the observability advantages of physical models and the unobservable behavior modeling capabilities of probabilistic models.
Data Source
AI summary
A machine learning system and method of operating a machine learning system for determining a time series, comprising providing an input for a first in particular generative model depending on a probabilistic variable, determining an output of the first model in response to the input for the first model, the output of the first model characterizing the time series. The first model comprises a first layer that is trained to map input for the first model determined depending on the probabilistic variable to output characterizing intermediate data, and a second layer that is trained to map the intermediate data to the time series depending on an output of a third layer of the first model. The output of the third layer characterizes a physical constraint to a machine state. Values of the time series or of the intermediate data are constrained by the output of the third layer.