Engine Neural Network Control With Recurrent Feature Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models for engine control face a reduction in generalization performance due to overlearning when the number of neurons or layers is increased, leading to poor prediction precision for unknown data.
Innovation Solution
Incorporating a fully connected layer between the input layer and the layer with a return path in the neural network, utilizing a rectified linear unit activation function to remove unnecessary features and compress data, thereby reducing the number of neurons and layers on the return path, which suppresses overlearning and improves prediction precision without increasing the degree of freedom.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of neurons or layers in the neural network is increased to improve prediction precision, then the model can capture more complex patterns, but overlearning occurs and generalization performance deteriorates
Solution Approach 1:
A fully connected layer is introduced as an intermediary component between the input layer and the layer with return path. This intermediate layer processes and compresses input features before they reach the recurrent structure, enabling the network to learn effective representations without requiring an excessive number of neurons or layers in the return path, thus maintaining generalization performance while achieving high prediction precision.
Solution Approach 2:
The patent changes the architectural parameters of the neural network by introducing a fully connected layer with a specific number of neurons (e.g., 10 neurons) and using rectified linear unit activation functions. This parameter configuration allows the network to achieve high prediction precision with a controlled degree of freedom, preventing overlearning while capturing complex temporal patterns in engine data.
2Measurement precision
If the number of neurons and layers on the return path is increased to improve prediction precision, then the model captures more temporal dependencies, but the degree of freedom increases leading to overlearning
Solution Approach 1:
The neural network architecture is segmented into distinct functional components: an input layer, a fully connected layer for feature extraction and compression, and a layer with return path for temporal processing. This segmentation allows each component to have a specialized and optimized size, reducing the overall degree of freedom while maintaining high prediction precision through coordinated operation of the segmented structures.
Solution Approach 2:
The patent extracts the feature extraction and compression function into a separate fully connected layer, removing the burden from the return path layer. This extraction allows the return path layer to focus solely on temporal dependency modeling with a manageable number of neurons, preventing overlearning while maintaining high prediction precision through the specialized分工 of the extracted component.
3Measurement precision
If the degree of freedom of the neural network is increased to improve prediction precision, then the model can learn more complex patterns, but learning specific to training data occurs reducing generalization
Solution Approach 1:
The patent adds a new dimensional aspect to the network architecture by introducing a fully connected layer that operates in a different functional dimension from the recurrent layer. This dimensional addition allows the network to process and compress features in one dimension while temporal processing occurs in another, enabling complex pattern learning with controlled total degrees of freedom, thus maintaining generalization capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An engine control device includes an engine model having a neural network that inputs a manipulated variable of the engine and computes a controlled variable; and a controller that computes the manipulated variable so as to reduce a deviation between the controlled variable and a target controlled variable. The neural network includes an input layer to which the manipulated variable are input; a first hidden layer including a first fully connected layer; a second hidden later including a second fully connected layer that generates a plurality of second output values at a first time and has a return path on which the plurality of second output values at a second time, earlier than the first time, are input into the second fully connected layer; and an output layer from which the plurality of second output values at the first time are output as the controlled variable.