Engine Control Neural Network with Multi-Input Parameter Combinations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network models for internal combustion engines face challenges in improving calculation accuracy of output parameters due to limited types of input parameters, as not all input parameters can be acquired, such as injection density and injection spread, which affect combustion results.

Innovation Solution

The implementation of a neural network model with multiple input layers, where different combinations of input parameters are input to each neural network unit, increasing the total number of apparent input parameters without increasing the types of input parameters, thereby generating diverse intermediate state quantities and enhancing the model's expressive power.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the number of input parameters is increased to improve calculation accuracy, then the information amount for calculating the output parameter increases, but the types of input parameters that can be acquired remain limited

Engineering Contradiction:
Improvecalculation accuracy of output parameterVSAvoidtypes of input parameters
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent divides the neural network into multiple input layers, where each input layer receives a different combination of input parameters. This segmentation allows the system to process limited input parameter types through multiple pathways, effectively increasing the information amount without requiring additional parameter types.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from a single-input-layer architecture to a multi-input-layer architecture, adding a dimensional aspect to the neural network structure. This dimensional change enables the model to process the same limited input parameters in different combinations across multiple layers, thereby improving calculation accuracy without increasing parameter type diversity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If multiple input layers are added to increase the number of apparent input parameters, then the expressive power of the model is enhanced, but the device complexity increases

Engineering Contradiction:
Improvecalculation accuracy of output parameterVSAvoidneural network structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the neural network into multiple input layers with different parameter combinations, allowing each layer to specialize in processing specific parameter sets. This segmentation enhances expressive power while managing complexity by organizing the network structure in a systematic, modular way.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3690558B1Control device of internal combustion engine, in-vehicle electronic control unit, machine learning system, control method of internal combustion engine, manufacturing method of electronic control unit, and output parameter calculation device
Publication Date: 2022.05.04 TOYOTA JIDOSHA KK
  • EP3690558B1 patent drawingFigure 1
  • EP3690558B1 patent drawingFigure 2~3
  • EP3690558B1 patent drawingFigure 4

AI summary

A control device of an internal combustion engine includes a parameter acquisition unit (81) that acquires a plurality of input parameters, a calculation unit (82) that calculates at least one output parameter using a neural network model, and a controller (83) that controls the internal combustion engine. The neural network model includes a plurality of neural network units and an output layer. Each of the neural network units includes one input layer and at least one intermediate layer. The neural network model inputs different combinations of input parameters selected from the input parameters to each of the input layers of the neural network units such that a total number of input parameters to be input to the neural network units is larger than the number of the input parameters.