Powertrain Quantity Estimation With Likelihood-Based Model Switching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for estimating quantities in a motor vehicle powertrain, such as temperature or flow rates, using neural networks face gaps in coverage during supervised learning, leading to potential divergence when encountering conditions not covered by the training plan, necessitating improved accuracy and computing efficiency.
Innovation Solution
A system that combines a supervised-learning-based estimating module with a likelihood-checking module and a downstream processing module, switching between neural network outputs and a physical model based on a likelihood index, ensuring accurate and efficient estimation by using a fallback physical model when conditions deviate from the training domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a supervised-learning-based neural network is used to estimate quantities, then estimation accuracy is improved for training-covered conditions, but reliability deteriorates when encountering conditions not covered by the training plan
Solution Approach 1:
A likelihood-checking module is introduced as an intermediary between the neural network estimator and the final output. This module evaluates whether current operating conditions fall within the training domain by comparing input data against stored reference data, and only allows the neural network output to pass through when the likelihood threshold is satisfied, thereby ensuring reliability without compromising accuracy
2Reliability
If a physical model is used as fallback, then reliability is improved for uncovered conditions, but manufacturing precision deteriorates due to switching between models
Solution Approach 1:
The system implements feedback by continuously monitoring the likelihood index and dynamically switching between the neural network estimator and the physical model based on whether current conditions are covered by the training plan. This feedback mechanism ensures that the more accurate neural network is used when appropriate, while falling back to the physical model only when necessary, thus minimizing precision loss
3Reliability
If a likelihood-checking module is added, then output reliability is improved, but device complexity increases
Solution Approach 1:
The likelihood-checking module is designed to perform multiple functions: it evaluates the likelihood of current conditions, compares input data against reference data, determines whether to use the neural network or physical model, and manages the switching logic. By consolidating these functions into a single multi-functional module, the increase in system complexity is minimized while achieving the reliability improvement
Data Source
AI summary
A method and system for determining iteratively a quantity of interest used to control at least one component of a motor-vehicle powertrain. The system includes a supervised-learning-based estimating module, for example a neural network, which receives as input a set of first input data, and delivers as output at least one intermediate output quantity, a likelihood-checking module, which receives as input the set of first input data and a set of second input data, and which delivers as output a likelihood index, a downstream processing module which receives as input the likelihood index and the intermediate output quantity, and which delivers as output a final output quantity. When the likelihood index is good, the final output quantity is obtained from the intermediate output quantity, otherwise the final output quantity is obtained from a model of a physical representation.