Missing Sensor Data Classification via Conditional Normalizing Flow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for reconstructing missing sensor data often result in inaccurate environmental state assessments due to their reliance solely on sensor data, neglecting the non-linear relationship between environmental states and sensor values.
Innovation Solution
A method utilizing machine learning to generate multiple hypotheses for missing sensor values, where the output of the technical system is used to ascertain an expected value based on these hypotheses, treating the missing sensor value as a random variable and using a conditional normalizing flow to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods are used to reconstruct missing sensor values based solely on other sensor values, then the process is simple, but the accuracy of environmental state determination deteriorates due to non-linear relationships
Solution Approach 1:
The patent applies dynamics by treating the missing sensor value as a random variable rather than a fixed value, allowing the system to consider multiple possible values and their probabilities. This dynamic approach enables the system to adapt to the non-linear relationship between environmental states and sensor values, improving measurement precision while managing complexity through probabilistic modeling.
Solution Approach 2:
The patent changes the parameter representation by introducing a probability distribution over possible sensor values. Instead of working with single deterministic values, the system works with probability distributions that capture the uncertainty and non-linear relationships, thereby improving accuracy of environmental state determination.
2Measurement precision
If multiple hypotheses are generated for missing sensor values using machine learning, then the accuracy of environmental state determination improves, but the computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning system (conditional normalizing flow) to learn the relationships between sensor values and environmental states. This pre-training allows the system to quickly generate accurate hypotheses for missing values during operation, improving classification and regression accuracy while managing computational complexity through efficient inference.
Solution Approach 2:
The patent introduces a conditional normalizing flow as an intermediary between the input sensor values and the output hypotheses. This intermediary model learns the complex non-linear relationships and transforms the input data into a form that makes hypothesis generation more efficient and accurate, thereby improving measurement precision while controlling the complexity of the overall system.
Data Source
AI summary
A computer-implemented method for ascertaining a classification and/or a regression result based on the plurality of sensor values. The method includes: ascertaining a plurality of hypotheses regarding a missing sensor value using a machine learning system; ascertaining a plurality of outputs, an output being based in each case on the plurality of sensor values and a hypothesis and the output characterizing a classification and/or a regression result; providing an aggregation of the plurality of outputs as the classification and/or the regression result.