Real-time prediction method for engine emission

Patent metadata

Field	Value
Publication number	US-12450492-B2
Application number	US-202217674762-A
Country	US
Kind code	B2
Filing date	Feb 17, 2022
Priority date	Feb 18, 2021
Publication date	Oct 21, 2025
Grant date	Oct 21, 2025

Abstract

A real-time prediction method for an engine emission is provided, including: acquiring multiple known historical test data samples for engine emission, dividing the samples into a training set and a test set to train multiple neural network (NN), calculating mean square errors (MSEs) output by each of the NNs with the different numbers of hidden layer nodes to determine a topological structure of the NNs, optimizing initial weights and initial thresholds for each of the NNs with a mind evolutionary algorithm (MEA), and establishing a real-time engine emission prediction system with an Adaboost algorithm.

First claim

What is claimed is:

1. A real-time prediction method for an engine emission, comprising following steps:
building a real-time prediction device comprising one or more processors and a strong neural network (NN) predictor model, comprising:
step 1: acquiring a plurality of known historical test data samples for engine emission, the historical test data samples each comprising values of evaluation indexes affecting the engine emission and an emission value of an engine in a working condition, wherein the evaluation indexes affecting the engine emission comprises a rotational speed, a torque, a power, a rail pressure, an air-fuel ratio, a fuel consumption, an exhaust gas recirculation (EGR) rate and a start of injection (SOI), and the emission value of the engine in the working condition comprises an nitrogen oxide (NOx) emission value, a total hydro carbon (THC) emission value and a carbon monoxide (CO) emission value;
step 2: randomly dividing the plurality of known historical test data samples into a training set and a test set;
step 3: determining, for each of a plurality of (NN) predictive models, a topological structure comprising a number of input layer nodes, a number of output layer nodes and a number of hidden layer nodes, comprising:
step 31: determining, with values of evaluation indexes affecting the engine emission in the training set for the historical test data samples as an input of each of NNs, and an emission value of the engine in the working condition in the training set for the historical test data samples as an output of each of the NNs, the number of input layer nodes and the number of output layer nodes in the NN according to input and output dimensions, setting a number of hidden layer nodes in each of the plurality of NN predictive models to be different, and training the plurality of NN predictive models;
step 32: obtaining a plurality of trained NN predictive models upon completion of the training on the NN predictive models, and respectively calculating, with values of evaluation index affecting the engine emission in the test set for the historical test data samples as an input of each of the trained NNs, mean square errors (MSEs) between predicted values of the trained NNs and theoretical values; and
step 33: setting a number of hidden layer nodes corresponding to a minimum MSE of the MSEs between the predicted values and the theoretical value of the trained NN predictive models as an optimal number of hidden layer nodes, and setting the number of hidden layer nodes in each of rest trained NN predictive models as the optimal number of hidden layer nodes, to obtain a plurality of identical NN predictive models having identical topological structure;

step 4: optimizing, with a mind evolutionary algorithm (MEA), randomly generated winning subpopulations and temporary subpopulations each containing individuals, to obtain optimal initial weights and optimal initial thresholds for each of the plurality of identical NN predictive models, wherein each individual is decoded to obtain initial weights and initial thresholds for each of the plurality of identical NN predictive models; and
step 5: training, with the optimal initial weights and the optimal initial thresholds obtained with the MEA as the initial weights and the initial thresholds for each of the plurality of identical NN predictive models, each of the identical NN predictive models with the training set for the historical test data samples through an Adaboost algorithm, to obtain a plurality of weak NN predictors, and combining the plurality of weak NN predictors into the strong NN predictor model; and

receiving, by the real-time prediction device, values of evaluation indexes affecting the engine emission which are measured in real time on the engine, loading the values to the strong NN predictor model, and outputting, by the real-time prediction device, a real-time predicted result for the engine emission in real time, wherein the real-time predicted result comprises a real-time NOx emission value, a real-time THC emission value, and a real-time CO emission value.

2. The real-time prediction method for the engine emission according to claim 1 , wherein the training in step 31 of step 3 specifically comprises:
step 301: performing min-max normalization on an input k and an output 1 of each sample i in the training set for the historical test data samples according to equations:

x
k
i

=

x
k
i

-

x

k
⁢

min

x

k
⁢

max

-

x

k
⁢

min

⁢

i
=
1

,
2
,
…

,

m
;

k
=
1

,
2
,
…

,
p

(

3
-
1

)

y
l
i

=

y
l
i

-

y

l
⁢

min

Assignees

• Univ Tianjin

Inventors

• Liu Haifeng
• Zhang Xiaoteng
• Wang Can
• Yao Mingfa

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G01M15/102

  by monitoring exhaust gases · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title